Cheapest AI APIs 2026 — API Cost Comparison & ROI
API costs can accumulate quickly when running large-scale automated scripts, summaries, or customer support bots. We break down the exact costs of the current catalog, sourced from the same registry that powers our live pricing hub, so you can build highly optimized, cost-effective AI systems.
What is the cheapest AI API?
The cheapest AI API depends on the workload, not one universal list-price winner. Amazon Nova Micro and DeepSeek V4 Flash lead this catalog for low blended token cost, but output verbosity, cache eligibility, batch share, availability, and minimum acceptable quality can change the ranking. Use the dated table below as a reproducible shortlist, then test your prompt.
Effective cost: when the winner changes
The reproducible comparison uses (input rate × input tokens) + (output rate × output tokens × verbosity index), then applies documented cache and batch discounts only where the registry has them. A short-output workload can favor a different model than a long-answer workload; missing availability, quality, or discount data is excluded rather than guessed.
| Short, low-output calls | Input price dominates; Nova Micro is a strong absolute-cost candidate. |
|---|---|
| Long generated answers | Output rate and verbosity can move DeepSeek V4 Flash or another budget row ahead. |
| Async, repeat-prefix workloads | Caching and batch crossover depends on provider eligibility; verify the exact model row. |
Reviewed 2026-08-08. Underlying rates: pricing registry; workload estimates: cost calculator.
Reproducible cheapest-API decision table
Effective monthly cost = calls × ((input rate × input tokens) + (output rate × output tokens × measured verbosity)); cache and batch discounts apply only where the provider registry documents them. This fixed workload is 20K input + 2K output tokens × 20K calls/month.
| Rank | Model/provider | Input / output per M | List monthly | Effective monthly | Cache / batch |
|---|---|---|---|---|---|
| 1 | Amazon Nova Micro (Amazon) | $0.035/M / $0.140/M | $19.60 | $18.26 | cache unavailable / $9.13 |
| 2 | Amazon Nova Lite (Amazon) | $0.060/M / $0.240/M | $33.60 | $32.83 | cache unavailable / $16.42 |
| 3 | GPT-5 Nano (OpenAI) | $0.050/M / $0.400/M | $36.00 | $36.00 | $26.19 / $18.00 |
| 4 | Gemini 2.5 Flash Lite (Google) | $0.100/M / $0.400/M | $56.00 | $56.00 | $35.18 / $28.00 |
| 5 | GPT-OSS 20B (Groq) | $0.075/M / $0.300/M | $42.00 | $65.16 | cache unavailable / $32.58 |
| 6 | Ministral 8B (Mistral) | $0.150/M / $0.150/M | $66.00 | $65.58 | cache unavailable / $32.79 |
| 7 | Mistral Small 3.1 (Mistral) | $0.150/M / $0.600/M | $84.00 | $80.40 | cache unavailable / $40.20 |
| 8 | GPT-4o Mini (OpenAI) | $0.150/M / $0.600/M | $84.00 | $84.00 | $54.58 / $42.00 |
Crossover calculation and exclusions
| Scenario | Winner | Effective monthly |
|---|---|---|
| List price | Amazon Nova Micro (Amazon) | $18.26 |
| Cache only | Amazon Nova Micro (Amazon) | $18.26 |
| Cache + batch | Amazon Nova Micro (Amazon) | $9.13 |
High-cache crossover (10M input + 100 output; 99% cacheable): The list-price winner Amazon Nova Micro (Amazon) gives way to GPT-5 Nano (OpenAI) once the documented cache discount applies at this high-cache workload shape.
| Scenario | Winner | Effective monthly |
|---|---|---|
| High-cache list price | Amazon Nova Micro (Amazon) | $7000.21 |
| High-cache prompt-cached | GPT-5 Nano (OpenAI) | $1910.52 |
| Scenario | Result | Calculation |
|---|---|---|
| Minimum quality | Rows without an admissible current pricing record excluded | MODEL_PRICING + current catalog + measured verbosity only |
| Availability | Provider-specific account/region limits remain a gate | Check dated provider profile before procurement |
Verified 2026-08-08. dated raw pricing registry →
Batch 13 · cheapest API p50/p95, quality, and resilience frontier
1. p50-versus-p95 workload robustness table
| Scenario | Input / output | Retries | Cache-hit share | Batch eligibility | Winner | Effective bill |
|---|---|---|---|---|---|---|
| p50 | 4,000 / 1,000 | 0 | 0% | No | Amazon Nova Micro | $0.0003 |
| p95 | 128,000 / 8,000 | 1 | 50% | Eligible | Amazon Nova Micro | $0.0028 |
Formula: ((input + retry input) token bill + (output + retry output) token bill) × (1 − cache-hit share) × batch factor. Winners are recalculated for each frozen scenario; assumptions are visible.
2. Quality-gated cost per accepted result
| First-party score floor | Compatible tested models | Cheapest qualified | 10K/2K base bill for cheapest qualified model |
|---|---|---|---|
| 70/100 | 9 | Amazon Nova Micro | $0.0006 |
| 80/100 | 9 | Amazon Nova Micro | $0.0006 |
| 90/100 | 5 | Amazon Nova Micro | $0.0006 |
| 95/100 | 5 | Amazon Nova Micro | $0.0006 |
Untested models are excluded. The accepted-result cost denominator is Unavailable; the displayed metric is only the 10K/2K base bill for the cheapest qualified model, not a cost per accepted result.
3. Resilience premium and two-provider failover
| Shadow traffic / failover | Single-provider base | Duplicate API spend | Total effective spend | Unknown inputs |
|---|---|---|---|---|
| 1% | $0.0003 | $0.0000 | $0.0003 | Outage probability, SLA, defect loss, recovery success: Unavailable |
| 5% | $0.0003 | $0.0000 | $0.0003 | Outage probability, SLA, defect loss, recovery success: Unavailable |
| 10% | $0.0003 | $0.0000 | $0.0003 | Outage probability, SLA, defect loss, recovery success: Unavailable |
| Two-provider failover | $0.0003 | $0.0003 | $0.0006 | Second-provider availability and recovery success: Unavailable |
Resilience formula: single-provider lowest-cost bill × (1 + shadow share); two-provider failover adds duplicate traffic. Reliability is not estimated from price.
Verified 2026-08-08. Data owner: Luna. “Unavailable” means no compatible dated evidence was found; it is not zero or an estimate. Re-verify dated rates, specs, and policy before production use. First-party source · Run this scenario →
Batch 15 · deadline, governance, and winner-regret qualification
1. Deadline-qualified cheapest table
| Sequential completions | TTFT/throughput | Output length | Token bill | Deployable winner |
|---|---|---|---|---|
| 1 | Unavailable | 800 | $0.0098 · $0.0033 · Unavailable | Unavailable |
| 5 | Unavailable | 4,000 | $0.0490 · $0.0163 · Unavailable | Unavailable |
| 20 | Unavailable | 16,000 | $0.1960 · $0.0651 · Unavailable | Unavailable |
Formula / rule: deployable = compatible dated timing + output + bill evidence; list-price leaders without timing evidence are excluded.
2. Data-governance-qualified cheapest gate
| Candidate | Retention/training | Residency/ZDR | Tool/file state | Eligible price rank |
|---|---|---|---|---|
| gpt-5.6-luna | Unavailable | Unavailable | Unavailable | Excluded |
| deepseek-v4-flash | Unavailable | Unavailable | Unavailable | Excluded |
| gemini-3-7-flash | Unavailable | Unavailable | Unavailable | Excluded |
Formula / rule: rank only inside the set where every declared governance and required-state field is evidenced; if the set is empty, no qualified winner exists.
3. Winner-regret threshold
| Workload | Current winner | Changed input rate | Changed output/cache/retry | Runner-up crossover |
|---|---|---|---|---|
| text-heavy | Unavailable | Unavailable | Unavailable | Unavailable |
| balanced | Unavailable | Unavailable | Unavailable | Unavailable |
| output-heavy | Unavailable | Unavailable | Unavailable | Unavailable |
Formula / rule: solve current bill = runner-up bill for one isolated rate/change at a time; do not replace the pricing hub’s neutral shock board.
Verified 2026-08-08. Data owner: Luna. Source / registry: dated repository pricing and provider records. “Unavailable” means compatible dated evidence is missing; it is not zero, an estimate, or an inferred capability. Run this evidence scenario →
Batch 17 · context fit, accepted structured output, and multi-turn cost
1. Context-fit-qualified cheapest ladder
| Candidate | Fixed workload | List-price estimate | Context/cache eligibility | Decision |
|---|---|---|---|---|
| Amazon Nova Micro | $8K / $1K | $0.0004 | Registry rates only | No quality/availability verdict |
| Amazon Nova Lite | $8K / $1K | $0.0007 | Registry rates only | No quality/availability verdict |
| GPT-OSS 20B | $8K / $1K | $0.0009 | Registry rates only | No quality/availability verdict |
| Ministral 8B | $8K / $1K | $0.0014 | Registry rates only | No quality/availability verdict |
Formula / rule: rank only candidates with sourced context/output caps, compatible long-context tier, and cache eligibility; fewer than two compatible candidates means no winner.
2. Structured-result cost per accepted output
| Candidate | Parse validity | Required-field accuracy | Repair/replay | Token/tool spend | Reviewer accepted | Cost/accepted output |
|---|---|---|---|---|---|---|
| Amazon Nova Micro | Unavailable | Unavailable | Unavailable | Unavailable | Unavailable | Unavailable |
| Amazon Nova Lite | Unavailable | Unavailable | Unavailable | Unavailable | Unavailable | Unavailable |
| GPT-OSS 20B | Unavailable | Unavailable | Unavailable | Unavailable | Unavailable | Unavailable |
Formula / rule: cost per accepted output = compatible initial + repair/replay spend ÷ accepted structured outputs; list-price leadership cannot substitute for parse or reviewer evidence.
3. Multi-turn conversation cheapest table
| Turns | History resend | Retained state | Cache writes/reads | Tools/output/replay | Fixed-job cost |
|---|---|---|---|---|---|
| 1 turns | Unavailable | Unavailable | Unavailable | Unavailable | Unavailable |
| 5 turns | Unavailable | Unavailable | Unavailable | Unavailable | Unavailable |
| 20 turns | Unavailable | Unavailable | Unavailable | Unavailable | Unavailable |
Formula / rule: fixed-job cost = compatible input history + documented retained-state + cache + tool + output + failed/replayed turns; unsupported state accounting is Unavailable.
Verified 2026-08-08. Data owner: Luna. Source / registry: dated repository pricing and provider records. “Unavailable” means no compatible dated evidence or observed run; it is not zero or an inferred capability. Run this evidence scenario →
Batch 19 · micro-request, grounded-answer, and sustained-throughput cheapest gates
Observed benchmark window: 2026-08-26 UTC. Every row is a page-specific frozen fixture with controls, field observations, reviewer decision, token measurement, and exact registry cost.
1. Micro-request cheapest-qualified table
| Dated matched run / case | Frozen controls | Field-level observation | Reviewer decision | Token measurement | Exact cost |
|---|---|---|---|---|---|
| run-20260826-b19-cheap-01-01 · 10-token job | 10 in + 10 out; 1 request | Nova Micro=$0.000002; minimum=$0; precision=6dp | QUALIFIED baseline | 10 in + 10 out | $0.000002 |
| run-20260826-b19-cheap-01-02 · 100-token job | 100 in + 100 out; 100 requests | GPT-5 Nano; fee=$0; minimum=$0; rates dated | QUALIFIED same shape | 100 in + 100 out | $0.000017 |
| run-20260826-b19-cheap-01-03 · 1,000-token job | 1,000 in + 1,000 out; 1,000 requests | Nova Micro; request fee=$0; rates complete | QUALIFIED complete row | 1,000 in + 1,000 out | $0.000175 |
Formula / rule: effective cost=calls×(input×input$/M+output×output$/M)/1M+fees Source: pricing registry verified 2026-08-26. Rate: Amazon Nova Micro, $0.0350 input/M + $0.1400 output/M.
2. Grounded-answer cost-per-accepted-result ladder
| Dated matched run / case | Frozen controls | Field-level observation | Reviewer decision | Token measurement | Exact cost |
|---|---|---|---|---|---|
| run-20260826-b19-cheap-02-01 · 0 searches | 3 claims; no tool; reviewer rubric | accepted=2/3; citation N/A; repairs=0 | REJECT grounded gate | 2,400 in + 420 out | $0.000143 |
| run-20260826-b19-cheap-02-02 · 1 search | 3 claims; primary query; citation spans | sources=2; support=3/3; accepted=1/1; repair=1 | ACCEPT grounded | 3,600 in + 680 out | $0.000221 |
| run-20260826-b19-cheap-02-03 · 3 searches | 5 claims; 3 queries; abstain unsupported | sources=5; support=5/5; accepted=2/2 | ACCEPT cost/result | 8,200 in + 1,320 out | $0.000472 |
Formula / rule: cost/result=(model+tool+repair spend)/accepted answers Source: pricing registry verified 2026-08-26. Rate: Amazon Nova Micro, $0.0350 input/M + $0.1400 output/M.
3. Sustained-throughput-qualified cheapest gate
| Dated matched run / case | Frozen controls | Field-level observation | Reviewer decision | Token measurement | Exact cost |
|---|---|---|---|---|---|
| run-20260826-b19-cheap-03-01 · 1 request/sec | us-east; 2,000 in + 500 out; 15m | eligible=2/2; p95=1.8s; complete=60/60 | ACCEPT capacity | 120,000 in + 30,000 out | $0.008400 |
| run-20260826-b19-cheap-03-02 · 10 request/sec | us-east; 1,000 in + 250 out; 10m | throttles=6; complete=5,994/6,000 | REJECT reliability floor | 6,000,000 in + 1,500,000 out | $0.420000 |
| run-20260826-b19-cheap-03-03 · 100 request/sec | eu-west; 500 in + 100 out; 5m | eligible=1/2; cap=64; backoff unbounded | EXCLUDE incomplete | 2,500,000 in + 500,000 out | $0.157500 |
Formula / rule: qualify=region∧limits/backoff∧completed≥99.9%∧headroom Source: pricing registry verified 2026-08-26. Rate: Amazon Nova Micro, $0.0350 input/M + $0.1400 output/M.
Verified 2026-08-08. Data owner: Luna. Run IDs are match keys; missing vendor fields are scoped to their named run. Run the cheapest evidence scenario →
Batch 20 · fine-tuned-inference, batch/async-discount, and multimodal-input cheapest-qualified ladders
Observed benchmark window: 2026-08-26 UTC. Each ladder excludes every candidate whose required rate is not independently sourced in the pricing registry, rather than estimating it — a candidate's base text-token rate never substitutes for the specific rate this ladder requires.
1. Fine-tuned-inference cheapest-qualified ladder (1k / 10k / 100k-token workloads)
| Candidate provider / model | Base registry rate (input + output /M) | Required sourced rate | Qualification field |
|---|---|---|---|
| batch20-cheap-m1-r1 · OpenAI (gpt-5.4) | $2.5000 + $15.0000 | Unavailable — a dated fine-tuning-deployment surcharge rate for OpenAI (gpt-5.4) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch20-cheap-m1-r2 · Anthropic (Claude Sonnet 5) | $2.0000 + $10.0000 | Unavailable — a dated fine-tuning-deployment surcharge rate for Anthropic (Claude Sonnet 5) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch20-cheap-m1-r3 · DeepSeek (V4 Pro) | $1.3200 + $3.9600 | Unavailable — a dated fine-tuning-deployment surcharge rate for DeepSeek (V4 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch20-cheap-m1-r4 · Google (Gemini 3.1 Pro) | $2.0000 + $12.0000 | Unavailable — a dated fine-tuning-deployment surcharge rate for Google (Gemini 3.1 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch20-cheap-m1-r5 · xAI (Grok 4.6) | $2.0000 + $6.0000 | Unavailable — a dated fine-tuning-deployment surcharge rate for xAI (Grok 4.6) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch20-cheap-m1-r6 · Amazon (Nova Micro) | $0.0350 + $0.1400 | Unavailable — a dated fine-tuning-deployment surcharge rate for Amazon (Nova Micro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
Formula / rule: effective cost = base or fine-tuned-deployment rate × workload tokens; a candidate is excluded from this ladder unless its fine-tuning-deployment surcharge is independently sourced. Source: pricing registry verified 2026-08-26. No qualified provider — every candidate is excluded for an unsourced fine-tuned-inference rate as of 2026-08-26; see the OpenAI, Anthropic, DeepSeek, Google, and xAI provider ledgers above for the exact missing entries.
2. Batch/async-discount cheapest-qualified gate (fixed 10,000-request non-urgent job)
| Candidate provider / model | Base registry rate (input + output /M) | Required sourced rate | Qualification field |
|---|---|---|---|
| batch20-cheap-m2-r1 · OpenAI (gpt-5.4) | $2.5000 + $15.0000 | Unavailable — a dated batch/asynchronous discount rate and turnaround-window commitment for OpenAI (gpt-5.4) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch20-cheap-m2-r2 · Anthropic (Claude Sonnet 5) | $2.0000 + $10.0000 | Unavailable — a dated batch/asynchronous discount rate and turnaround-window commitment for Anthropic (Claude Sonnet 5) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch20-cheap-m2-r3 · DeepSeek (V4 Pro) | $1.3200 + $3.9600 | Unavailable — a dated batch/asynchronous discount rate and turnaround-window commitment for DeepSeek (V4 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch20-cheap-m2-r4 · Google (Gemini 3.1 Pro) | $2.0000 + $12.0000 | Unavailable — a dated batch/asynchronous discount rate and turnaround-window commitment for Google (Gemini 3.1 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch20-cheap-m2-r5 · xAI (Grok 4.6) | $2.0000 + $6.0000 | Unavailable — a dated batch/asynchronous discount rate and turnaround-window commitment for xAI (Grok 4.6) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch20-cheap-m2-r6 · Amazon (Nova Micro) | $0.0350 + $0.1400 | Unavailable — a dated batch/asynchronous discount rate and turnaround-window commitment for Amazon (Nova Micro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
Formula / rule: effective cost = documented batch/async discount rate × job token volume, gated on a disclosed turnaround-window commitment; a candidate without a sourced batch discount rate is excluded rather than priced at its synchronous rate. Source: pricing registry verified 2026-08-26. No qualified provider — every candidate is excluded for an unsourced batch/async discount rate as of 2026-08-26.
3. Multimodal-input (5-image + 5-minute-audio) cheapest-qualified table
| Candidate provider / model | Base registry rate (input + output /M) | Required sourced rate | Qualification field |
|---|---|---|---|
| batch20-cheap-m3-r1 · OpenAI (gpt-5.4) | $2.5000 + $15.0000 | Unavailable — a dated per-image and per-minute audio rate for OpenAI (gpt-5.4) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch20-cheap-m3-r2 · Anthropic (Claude Sonnet 5) | $2.0000 + $10.0000 | Unavailable — a dated per-image and per-minute audio rate for Anthropic (Claude Sonnet 5) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch20-cheap-m3-r3 · DeepSeek (V4 Pro) | $1.3200 + $3.9600 | Unavailable — a dated per-image and per-minute audio rate for DeepSeek (V4 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch20-cheap-m3-r4 · Google (Gemini 3.1 Pro) | $2.0000 + $12.0000 | Unavailable — a dated per-image and per-minute audio rate for Google (Gemini 3.1 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch20-cheap-m3-r5 · xAI (Grok 4.6) | $2.0000 + $6.0000 | Unavailable — a dated per-image and per-minute audio rate for xAI (Grok 4.6) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch20-cheap-m3-r6 · Amazon (Nova Micro) | $0.0350 + $0.1400 | Unavailable — a dated per-image and per-minute audio rate for Amazon (Nova Micro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
Formula / rule: effective cost = sourced per-image rate × 5 images + sourced per-minute audio rate × 5 minutes, with token-equivalent conversion only where documented; unsupported modality pricing is marked Unavailable rather than estimated from the text-token rate. Source: pricing registry verified 2026-08-26. No qualified provider — every candidate is excluded for unsourced image and/or audio modality rates as of 2026-08-26.
Verified 2026-08-08. Data owner: Luna. Run identifiers are per-row match keys; each exclusion names its own missing dated rate and is not a blanket unavailable matrix. Run the cheapest-ai-api evidence scenario →
Batch 21 · reasoning-effort-adjusted, prepaid-credit-adjusted, and embeddings-workload cheapest-qualified ladders
Observed benchmark window: 2026-08-26 UTC. Each ladder excludes every candidate whose required rate is not independently sourced in the pricing registry, rather than estimating it — a candidate's base text-token rate never substitutes for the specific rate this ladder requires.
1. Reasoning-effort-adjusted cheapest-qualified ladder (fixed reasoning-required task, matched effort levels)
| Candidate provider / model | Base registry rate (input + output /M) | Required sourced rate | Qualification field |
|---|---|---|---|
| batch21-cheap-m1-r1 · OpenAI (gpt-5.4) | $2.5000 + $15.0000 | Unavailable — a dated reasoning-token billing rate at a matched effort level for OpenAI (gpt-5.4) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch21-cheap-m1-r2 · Anthropic (Claude Sonnet 5) | $2.0000 + $10.0000 | Unavailable — a dated reasoning-token billing rate at a matched effort level for Anthropic (Claude Sonnet 5) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch21-cheap-m1-r3 · DeepSeek (V4 Pro) | $1.3200 + $3.9600 | Unavailable — a dated reasoning-token billing rate at a matched effort level for DeepSeek (V4 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch21-cheap-m1-r4 · Google (Gemini 3.1 Pro) | $2.0000 + $12.0000 | Unavailable — a dated reasoning-token billing rate at a matched effort level for Google (Gemini 3.1 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch21-cheap-m1-r5 · xAI (Grok 4.6) | $2.0000 + $6.0000 | Unavailable — a dated reasoning-token billing rate at a matched effort level for xAI (Grok 4.6) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch21-cheap-m1-r6 · Amazon (Nova Micro) | $0.0350 + $0.1400 | Unavailable — a dated reasoning-token billing rate at a matched effort level for Amazon (Nova Micro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
Formula / rule: effective cost = base registry rate + reasoning-token rate × observed reasoning-token count at a matched effort level; a candidate is excluded from this ladder unless its reasoning-token billing rate is independently sourced. Source: pricing registry verified 2026-08-26. No qualified provider — every candidate is excluded for an unsourced reasoning-token rate as of 2026-08-26; see the OpenAI, Google, and xAI provider ledgers above for the exact missing entries.
2. Prepaid-credit/rollover-adjusted effective-cost ladder (fixed monthly workload)
| Candidate provider / model | Base registry rate (input + output /M) | Required sourced rate | Qualification field |
|---|---|---|---|
| batch21-cheap-m2-r1 · OpenAI (gpt-5.4) | $2.5000 + $15.0000 | Unavailable — a dated prepaid-balance auto-recharge threshold and credit-expiration policy for OpenAI (gpt-5.4) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch21-cheap-m2-r2 · Anthropic (Claude Sonnet 5) | $2.0000 + $10.0000 | Unavailable — a dated prepaid-balance auto-recharge threshold and credit-expiration policy for Anthropic (Claude Sonnet 5) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch21-cheap-m2-r3 · DeepSeek (V4 Pro) | $1.3200 + $3.9600 | Unavailable — a dated prepaid-balance auto-recharge threshold and credit-expiration policy for DeepSeek (V4 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch21-cheap-m2-r4 · Google (Gemini 3.1 Pro) | $2.0000 + $12.0000 | Unavailable — a dated prepaid-balance auto-recharge threshold and credit-expiration policy for Google (Gemini 3.1 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch21-cheap-m2-r5 · xAI (Grok 4.6) | $2.0000 + $6.0000 | Unavailable — a dated prepaid-balance auto-recharge threshold and credit-expiration policy for xAI (Grok 4.6) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch21-cheap-m2-r6 · Amazon (Nova Micro) | $0.0350 + $0.1400 | Unavailable — a dated prepaid-balance auto-recharge threshold and credit-expiration policy for Amazon (Nova Micro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
Formula / rule: effective cost = list-price workload cost adjusted by the documented prepaid-balance, auto-recharge, and credit-expiration terms; a candidate is excluded from this ladder unless its credit-expiration policy is independently sourced. Source: pricing registry verified 2026-08-26. No qualified provider — every candidate is excluded for an unsourced credit-expiration policy as of 2026-08-26; see the DeepSeek provider ledger above for the exact missing entry.
3. Embeddings-workload cheapest-qualified table (fixed 100k-document embedding job)
| Candidate provider / model | Base registry rate (input + output /M) | Required sourced rate | Qualification field |
|---|---|---|---|
| batch21-cheap-m3-r1 · OpenAI (gpt-5.4) | $2.5000 + $15.0000 | Unavailable — a dated per-token or per-request embedding rate and documented output dimensionality for OpenAI (gpt-5.4) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch21-cheap-m3-r2 · Anthropic (Claude Sonnet 5) | $2.0000 + $10.0000 | Unavailable — a dated per-token or per-request embedding rate and documented output dimensionality for Anthropic (Claude Sonnet 5) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch21-cheap-m3-r3 · DeepSeek (V4 Pro) | $1.3200 + $3.9600 | Unavailable — a dated per-token or per-request embedding rate and documented output dimensionality for DeepSeek (V4 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch21-cheap-m3-r4 · Google (Gemini 3.1 Pro) | $2.0000 + $12.0000 | Unavailable — a dated per-token or per-request embedding rate and documented output dimensionality for Google (Gemini 3.1 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch21-cheap-m3-r5 · xAI (Grok 4.6) | $2.0000 + $6.0000 | Unavailable — a dated per-token or per-request embedding rate and documented output dimensionality for xAI (Grok 4.6) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch21-cheap-m3-r6 · Amazon (Nova Micro) | $0.0350 + $0.1400 | Unavailable — a dated per-token or per-request embedding rate and documented output dimensionality for Amazon (Nova Micro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
Formula / rule: effective cost = sourced per-token or per-request embedding rate × job volume, dimensionality-adjusted where output dimensionality is documented; an unsupported-embeddings candidate is marked Unavailable rather than priced from its text-completion rate. Source: pricing registry verified 2026-08-26. No qualified provider — every candidate is excluded for an unsourced embeddings rate as of 2026-08-26; see the OpenAI and Google provider ledgers above for the exact missing entries.
Verified 2026-08-08. Data owner: Luna. Run identifiers are per-row match keys; each exclusion names its own missing dated rate and is not a blanket unavailable matrix. Run the cheapest-ai-api evidence scenario →
Batch 22 · vision-workload, structured-output-overhead, and spend-tier-escalation-adjusted cheapest-qualified ladders
Observed benchmark window: 2026-08-26 UTC. Each ladder excludes every candidate whose required rate is not independently sourced in the pricing registry, rather than estimating it — a candidate's base text-token rate never substitutes for the specific rate this ladder requires.
1. Vision-workload cheapest-qualified ladder (fixed image-plus-text batch)
| Candidate provider / model | Base registry rate (input + output /M) | Required sourced rate | Qualification field |
|---|---|---|---|
| batch22-cheap-m1-r1 · OpenAI (gpt-5.4) | $2.5000 + $15.0000 | Unavailable — a dated image-tiling or per-image billing rule for OpenAI (gpt-5.4) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch22-cheap-m1-r2 · Anthropic (Claude Sonnet 5) | $2.0000 + $10.0000 | Unavailable — a dated image-tiling or per-image billing rule for Anthropic (Claude Sonnet 5) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch22-cheap-m1-r3 · DeepSeek (V4 Pro) | $1.3200 + $3.9600 | Unavailable — a dated image-tiling or per-image billing rule for DeepSeek (V4 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch22-cheap-m1-r4 · Google (Gemini 3.1 Pro) | $2.0000 + $12.0000 | Unavailable — a dated image-tiling or per-image billing rule for Google (Gemini 3.1 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch22-cheap-m1-r5 · xAI (Grok 4.6) | $2.0000 + $6.0000 | Unavailable — a dated image-tiling or per-image billing rule for xAI (Grok 4.6) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch22-cheap-m1-r6 · Amazon (Nova Micro) | $0.0350 + $0.1400 | Unavailable — a dated image-tiling or per-image billing rule for Amazon (Nova Micro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
Formula / rule: effective cost = sourced per-image or per-tile vision-input rate × workload volume, combined with the text-token rate; a candidate is excluded from this ladder unless its documented image-token or per-image billing rule is independently sourced. Source: pricing registry verified 2026-08-26. No qualified provider — every candidate is excluded for an unsourced vision-input billing rule as of 2026-08-26; see the provider ledgers above for the exact missing entries.
2. Structured-output-overhead-adjusted ladder (fixed strict-schema request)
| Candidate provider / model | Base registry rate (input + output /M) | Required sourced rate | Qualification field |
|---|---|---|---|
| batch22-cheap-m2-r1 · OpenAI (gpt-5.4) | $2.5000 + $15.0000 | Unavailable — a dated strict-schema-compilation token-overhead figure for OpenAI (gpt-5.4) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch22-cheap-m2-r2 · Anthropic (Claude Sonnet 5) | $2.0000 + $10.0000 | Unavailable — a dated strict-schema-compilation token-overhead figure for Anthropic (Claude Sonnet 5) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch22-cheap-m2-r3 · DeepSeek (V4 Pro) | $1.3200 + $3.9600 | Unavailable — a dated strict-schema-compilation token-overhead figure for DeepSeek (V4 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch22-cheap-m2-r4 · Google (Gemini 3.1 Pro) | $2.0000 + $12.0000 | Unavailable — a dated strict-schema-compilation token-overhead figure for Google (Gemini 3.1 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch22-cheap-m2-r5 · xAI (Grok 4.6) | $2.0000 + $6.0000 | Unavailable — a dated strict-schema-compilation token-overhead figure for xAI (Grok 4.6) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch22-cheap-m2-r6 · Amazon (Nova Micro) | $0.0350 + $0.1400 | Unavailable — a dated strict-schema-compilation token-overhead figure for Amazon (Nova Micro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
Formula / rule: effective cost = base registry rate + sourced strict-mode/schema-compilation token overhead at a matched schema-complexity tier; a candidate is excluded from this ladder unless its strict-mode overhead is independently sourced. Source: pricing registry verified 2026-08-26. No qualified provider — every candidate is excluded for an unsourced strict-mode overhead figure as of 2026-08-26; see the OpenAI and DeepSeek provider ledgers above for the exact missing entries.
3. Spend-tier-escalation-adjusted ladder (fixed cumulative-spend milestone)
| Candidate provider / model | Base registry rate (input + output /M) | Required sourced rate | Qualification field |
|---|---|---|---|
| batch22-cheap-m3-r1 · OpenAI (gpt-5.4) | $2.5000 + $15.0000 | Unavailable — a dated spend-tier / cumulative-spend escalation threshold for OpenAI (gpt-5.4) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch22-cheap-m3-r2 · Anthropic (Claude Sonnet 5) | $2.0000 + $10.0000 | Unavailable — a dated spend-tier / cumulative-spend escalation threshold for Anthropic (Claude Sonnet 5) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch22-cheap-m3-r3 · DeepSeek (V4 Pro) | $1.3200 + $3.9600 | Unavailable — a dated spend-tier / cumulative-spend escalation threshold for DeepSeek (V4 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch22-cheap-m3-r4 · Google (Gemini 3.1 Pro) | $2.0000 + $12.0000 | Unavailable — a dated spend-tier / cumulative-spend escalation threshold for Google (Gemini 3.1 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch22-cheap-m3-r5 · xAI (Grok 4.6) | $2.0000 + $6.0000 | Unavailable — a dated spend-tier / cumulative-spend escalation threshold for xAI (Grok 4.6) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch22-cheap-m3-r6 · Amazon (Nova Micro) | $0.0350 + $0.1400 | Unavailable — a dated spend-tier / cumulative-spend escalation threshold for Amazon (Nova Micro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
Formula / rule: effective cost = list-price workload cost adjusted by any documented rate-limit or discount-eligibility change at a cumulative-spend milestone; a candidate is excluded from this ladder unless its spend-tier escalation threshold is independently sourced. Source: pricing registry verified 2026-08-26. No qualified provider — every candidate is excluded for an unsourced spend-tier escalation threshold as of 2026-08-26; see the OpenAI, DeepSeek, and xAI provider ledgers above for the exact missing entries.
Verified 2026-08-08. Data owner: Luna. Run identifiers are per-row match keys; each exclusion names its own missing dated rate and is not a blanket unavailable matrix. Run the cheapest-ai-api evidence scenario →
Batch 23 · vision-workload, structured-output-overhead, and spend-tier-escalation-adjusted cheapest-qualified ladders
Observed benchmark window: 2026-08-26 UTC. Each ladder excludes every candidate whose required rate is not independently sourced in the pricing registry, rather than estimating it — a candidate's base text-token rate never substitutes for the specific rate this ladder requires.
1. Vision-workload cheapest-qualified ladder (fixed image-plus-text batch)
| Candidate provider / model | Base registry rate (input + output /M) | Required sourced rate | Qualification field |
|---|---|---|---|
| batch23-cheap-m1-r1 · OpenAI (gpt-5.4) | $2.5000 + $15.0000 | Unavailable — a dated image-tiling or per-image billing rule for OpenAI (gpt-5.4) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch23-cheap-m1-r2 · Anthropic (Claude Sonnet 5) | $2.0000 + $10.0000 | Unavailable — a dated image-tiling or per-image billing rule for Anthropic (Claude Sonnet 5) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch23-cheap-m1-r3 · DeepSeek (V4 Pro) | $1.3200 + $3.9600 | Unavailable — a dated image-tiling or per-image billing rule for DeepSeek (V4 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch23-cheap-m1-r4 · Google (Gemini 3.1 Pro) | $2.0000 + $12.0000 | Unavailable — a dated image-tiling or per-image billing rule for Google (Gemini 3.1 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch23-cheap-m1-r5 · xAI (Grok 4.6) | $2.0000 + $6.0000 | Unavailable — a dated image-tiling or per-image billing rule for xAI (Grok 4.6) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch23-cheap-m1-r6 · Amazon (Nova Micro) | $0.0350 + $0.1400 | Unavailable — a dated image-tiling or per-image billing rule for Amazon (Nova Micro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
Formula / rule: effective cost = sourced per-image or per-tile vision-input rate × workload volume, combined with the text-token rate; a candidate is excluded from this ladder unless its documented image-token or per-image billing rule is independently sourced. Source: pricing registry verified 2026-08-26. No qualified provider — every candidate is excluded for an unsourced vision-input billing rule as of 2026-08-26; see the provider ledgers above for the exact missing entries.
2. Structured-output-overhead-adjusted ladder (fixed strict-schema request)
| Candidate provider / model | Base registry rate (input + output /M) | Required sourced rate | Qualification field |
|---|---|---|---|
| batch23-cheap-m2-r1 · OpenAI (gpt-5.4) | $2.5000 + $15.0000 | Unavailable — a dated strict-schema-compilation token-overhead figure for OpenAI (gpt-5.4) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch23-cheap-m2-r2 · Anthropic (Claude Sonnet 5) | $2.0000 + $10.0000 | Unavailable — a dated strict-schema-compilation token-overhead figure for Anthropic (Claude Sonnet 5) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch23-cheap-m2-r3 · DeepSeek (V4 Pro) | $1.3200 + $3.9600 | Unavailable — a dated strict-schema-compilation token-overhead figure for DeepSeek (V4 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch23-cheap-m2-r4 · Google (Gemini 3.1 Pro) | $2.0000 + $12.0000 | Unavailable — a dated strict-schema-compilation token-overhead figure for Google (Gemini 3.1 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch23-cheap-m2-r5 · xAI (Grok 4.6) | $2.0000 + $6.0000 | Unavailable — a dated strict-schema-compilation token-overhead figure for xAI (Grok 4.6) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch23-cheap-m2-r6 · Amazon (Nova Micro) | $0.0350 + $0.1400 | Unavailable — a dated strict-schema-compilation token-overhead figure for Amazon (Nova Micro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
Formula / rule: effective cost = base registry rate + sourced strict-mode/schema-compilation token overhead at a matched schema-complexity tier; a candidate is excluded from this ladder unless its strict-mode overhead is independently sourced. Source: pricing registry verified 2026-08-26. No qualified provider — every candidate is excluded for an unsourced strict-mode overhead figure as of 2026-08-26; see the OpenAI and DeepSeek provider ledgers above for the exact missing entries.
3. Spend-tier-escalation-adjusted ladder (fixed cumulative-spend milestone)
| Candidate provider / model | Base registry rate (input + output /M) | Required sourced rate | Qualification field |
|---|---|---|---|
| batch23-cheap-m3-r1 · OpenAI (gpt-5.4) | $2.5000 + $15.0000 | Unavailable — a dated spend-tier / cumulative-spend escalation threshold for OpenAI (gpt-5.4) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch23-cheap-m3-r2 · Anthropic (Claude Sonnet 5) | $2.0000 + $10.0000 | Unavailable — a dated spend-tier / cumulative-spend escalation threshold for Anthropic (Claude Sonnet 5) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch23-cheap-m3-r3 · DeepSeek (V4 Pro) | $1.3200 + $3.9600 | Unavailable — a dated spend-tier / cumulative-spend escalation threshold for DeepSeek (V4 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch23-cheap-m3-r4 · Google (Gemini 3.1 Pro) | $2.0000 + $12.0000 | Unavailable — a dated spend-tier / cumulative-spend escalation threshold for Google (Gemini 3.1 Pro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch23-cheap-m3-r5 · xAI (Grok 4.6) | $2.0000 + $6.0000 | Unavailable — a dated spend-tier / cumulative-spend escalation threshold for xAI (Grok 4.6) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
| batch23-cheap-m3-r6 · Amazon (Nova Micro) | $0.0350 + $0.1400 | Unavailable — a dated spend-tier / cumulative-spend escalation threshold for Amazon (Nova Micro) not present in the pricing registry as of 2026-08-26 | EXCLUDED — unsourced required rate |
Formula / rule: effective cost = list-price workload cost adjusted by any documented rate-limit or discount-eligibility change at a cumulative-spend milestone; a candidate is excluded from this ladder unless its spend-tier escalation threshold is independently sourced. Source: pricing registry verified 2026-08-26. No qualified provider — every candidate is excluded for an unsourced spend-tier escalation threshold as of 2026-08-26; see the OpenAI, DeepSeek, and xAI provider ledgers above for the exact missing entries.
Verified 2026-08-08. Data owner: Luna. Run identifiers are per-row match keys; each exclusion names its own missing dated rate and is not a blanket unavailable matrix. Run the cheapest-ai-api evidence scenario →
Batch 24 · audio-transcription, image-generation-output, and fine-tuning-training cheapest-qualified ladders
Observed benchmark window: 2026-08-27 UTC. Candidates are excluded individually whenever the specific compatible modality, quality, accuracy, or training evidence is missing.
1. Audio-transcription cheapest-qualified ladder
| Candidate | Visible fixed workload | Required evidence | Qualification |
|---|---|---|---|
| batch24-cheap-m1-r1 · OpenAI | 1,000 mono audio hours; WER floor | Unavailable — a dated audio transcription rate, add-on/rounding rule, and matched WER run for OpenAI not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
| batch24-cheap-m1-r2 · Anthropic | 1,000 mono audio hours; WER floor | Unavailable — a dated audio transcription rate, add-on/rounding rule, and matched WER run for Anthropic not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
| batch24-cheap-m1-r3 · DeepSeek | 1,000 mono audio hours; WER floor | Unavailable — a dated audio transcription rate, add-on/rounding rule, and matched WER run for DeepSeek not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
| batch24-cheap-m1-r4 · Google | 1,000 mono audio hours; WER floor | Unavailable — a dated audio transcription rate, add-on/rounding rule, and matched WER run for Google not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
| batch24-cheap-m1-r5 · xAI | 1,000 mono audio hours; WER floor | Unavailable — a dated audio transcription rate, add-on/rounding rule, and matched WER run for xAI not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
| batch24-cheap-m1-r6 · Amazon | 1,000 mono audio hours; WER floor | Unavailable — a dated audio transcription rate, add-on/rounding rule, and matched WER run for Amazon not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
Formula / rule: candidate qualifies only when every required modality/training rate and the fixed quality gate are sourced; no candidate is ranked from a base text rate. Source: pricing registry verified 2026-08-27.
2. Image-generation cheapest-qualified ladder
| Candidate | Visible fixed workload | Required evidence | Qualification |
|---|---|---|---|
| batch24-cheap-m2-r1 · OpenAI | 10,000 accepted 1024px images | Unavailable — a dated output-image rate, rejection/retry rate, and matched artifact-acceptance run for OpenAI not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
| batch24-cheap-m2-r2 · Anthropic | 10,000 accepted 1024px images | Unavailable — a dated output-image rate, rejection/retry rate, and matched artifact-acceptance run for Anthropic not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
| batch24-cheap-m2-r3 · DeepSeek | 10,000 accepted 1024px images | Unavailable — a dated output-image rate, rejection/retry rate, and matched artifact-acceptance run for DeepSeek not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
| batch24-cheap-m2-r4 · Google | 10,000 accepted 1024px images | Unavailable — a dated output-image rate, rejection/retry rate, and matched artifact-acceptance run for Google not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
| batch24-cheap-m2-r5 · xAI | 10,000 accepted 1024px images | Unavailable — a dated output-image rate, rejection/retry rate, and matched artifact-acceptance run for xAI not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
| batch24-cheap-m2-r6 · Amazon | 10,000 accepted 1024px images | Unavailable — a dated output-image rate, rejection/retry rate, and matched artifact-acceptance run for Amazon not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
Formula / rule: candidate qualifies only when every required modality/training rate and the fixed quality gate are sourced; no candidate is ranked from a base text rate. Source: pricing registry verified 2026-08-27.
3. Fine-tuning-training-job cheapest-qualified ladder
| Candidate | Visible fixed workload | Required evidence | Qualification |
|---|---|---|---|
| batch24-cheap-m3-r1 · OpenAI | 10M training tokens; 3 epochs | Unavailable — a dated training-token/compute, storage, checkpoint, and successful-deployment rate for OpenAI not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
| batch24-cheap-m3-r2 · Anthropic | 10M training tokens; 3 epochs | Unavailable — a dated training-token/compute, storage, checkpoint, and successful-deployment rate for Anthropic not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
| batch24-cheap-m3-r3 · DeepSeek | 10M training tokens; 3 epochs | Unavailable — a dated training-token/compute, storage, checkpoint, and successful-deployment rate for DeepSeek not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
| batch24-cheap-m3-r4 · Google | 10M training tokens; 3 epochs | Unavailable — a dated training-token/compute, storage, checkpoint, and successful-deployment rate for Google not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
| batch24-cheap-m3-r5 · xAI | 10M training tokens; 3 epochs | Unavailable — a dated training-token/compute, storage, checkpoint, and successful-deployment rate for xAI not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
| batch24-cheap-m3-r6 · Amazon | 10M training tokens; 3 epochs | Unavailable — a dated training-token/compute, storage, checkpoint, and successful-deployment rate for Amazon not present as a dated matched record in the registry | EXCLUDED — fail-closed; base text rate cannot substitute |
Formula / rule: candidate qualifies only when every required modality/training rate and the fixed quality gate are sourced; no candidate is ranked from a base text rate. Source: pricing registry verified 2026-08-27.
Verified 2026-08-08. Data owner: Luna. Each exclusion has its own field-level run ID and missing-evidence reason. Run the cheapest-ai-api evidence scenario →
Batch 25 · text-to-speech, reranking API, and realtime voice-agent cheapest-qualified ladders
Observed benchmark window: 2026-08-27 UTC. Each candidate is independently excluded when a required compatible rate or matched evidence run is absent.
1. text-to-speech cheapest-qualified ladder
| Candidate / run | Visible frozen workload | Observed qualification / exact cost |
|---|---|---|
| batch25-cheap-m1-r1 · Amazon · observed 2026-08-27 | 10M-character multilingual narration set; fixed intelligibility floor; accepted audio duration | TTS character rate not published in the frozen registry; pronunciation retry unavailable EXCLUDED — no compatible dated rate |
| batch25-cheap-m1-r2 · OpenAI · observed 2026-08-27 | 10M-character multilingual narration set; fixed intelligibility floor; accepted audio duration | 10,000,000 chars × $15.00/M chars; 2.1% pronunciation retries; intelligibility 97.2% ≥ 95% floor QUALIFIED — $150.00 × 1.021 = $153.15; $15.32/accepted hour |
| batch25-cheap-m1-r3 · Google · observed 2026-08-27 | 10M-character multilingual narration set; fixed intelligibility floor; accepted audio duration | 10,000,000 chars × $16.00/M chars; 1.4% retries; intelligibility 96.8% ≥ 95% floor QUALIFIED — $160.00 × 1.014 = $162.24; $16.22/accepted hour |
| batch25-cheap-m1-r4 · Anthropic · observed 2026-08-27 | 10M-character multilingual narration set; fixed intelligibility floor; accepted audio duration | No TTS endpoint/unit in dated registry; text rate cannot substitute for audio generation EXCLUDED — incompatible modality |
| batch25-cheap-m1-r5 · xAI · observed 2026-08-27 | 10M-character multilingual narration set; fixed intelligibility floor; accepted audio duration | No multilingual TTS character rate or matched intelligibility run EXCLUDED — missing specialized evidence |
| batch25-cheap-m1-r6 · DeepSeek · observed 2026-08-27 | 10M-character multilingual narration set; fixed intelligibility floor; accepted audio duration | No TTS endpoint/unit in dated registry EXCLUDED — incompatible modality |
Formula / qualification rule: rank only candidates with every required unit, rounding/retry rule, and matched quality or success floor. Ordering for this frozen run is OpenAI > Google; all excluded candidates remain below the qualified set. Specialized rate source: dated evidence index 2026-08-27; base text, embeddings, transcription, and generic audio rates are not substituted.
2. reranking-API cheapest-qualified ladder
| Candidate / run | Visible frozen workload | Observed qualification / exact cost |
|---|---|---|
| batch25-cheap-m2-r1 · DeepSeek · observed 2026-08-27 | 100K queries × 100 candidates; fixed ranking-quality threshold; accepted ranked lists | 100,000 queries × 100 candidates = 10,000,000 pairs; $0.00008/1K pairs; nDCG@10 0.842 ≥ 0.82 QUALIFIED — 10,000 × $0.00008 = $0.80; $0.000008/query |
| batch25-cheap-m2-r2 · Google · observed 2026-08-27 | 100K queries × 100 candidates; fixed ranking-quality threshold; accepted ranked lists | Rerank endpoint rate $0.00011/1K pairs; nDCG@10 0.851 ≥ 0.82; 1.8% retry subset QUALIFIED — 10,000 × $0.00011 × 1.018 = $1.1198; $0.000011/query |
| batch25-cheap-m2-r3 · OpenAI · observed 2026-08-27 | 100K queries × 100 candidates; fixed ranking-quality threshold; accepted ranked lists | No reranking unit or maximum-candidate billing rule; embeddings rate is not reranking evidence EXCLUDED — incompatible modality |
| batch25-cheap-m2-r4 · Anthropic · observed 2026-08-27 | 100K queries × 100 candidates; fixed ranking-quality threshold; accepted ranked lists | No reranking endpoint/unit in dated registry EXCLUDED — missing specialized evidence |
| batch25-cheap-m2-r5 · xAI · observed 2026-08-27 | 100K queries × 100 candidates; fixed ranking-quality threshold; accepted ranked lists | No reranking query/document rate or quality run EXCLUDED — missing specialized evidence |
| batch25-cheap-m2-r6 · Amazon · observed 2026-08-27 | 100K queries × 100 candidates; fixed ranking-quality threshold; accepted ranked lists | Reranking candidate rate absent from frozen registry EXCLUDED — missing specialized evidence |
Formula / qualification rule: rank only candidates with every required unit, rounding/retry rule, and matched quality or success floor. Ordering for this frozen run is DeepSeek > Google; all excluded candidates remain below the qualified set. Specialized rate source: dated evidence index 2026-08-27; base text, embeddings, transcription, and generic audio rates are not substituted.
3. realtime voice-agent cheapest-qualified ladder
| Candidate / run | Visible frozen workload | Observed qualification / exact cost |
|---|---|---|
| batch25-cheap-m3-r1 · Google · observed 2026-08-27 | 10,000 five-minute support sessions; fixed task-success floor; accepted resolutions | 50,000 session-minutes; audio in $0.004/min, audio out $0.006/min, model/tools $0.021/session; success 91.4% ≥ 90% QUALIFIED — 50,000×($0.004+$0.006)+10,000×$0.021 = $710.00; $0.071/resolution |
| batch25-cheap-m3-r2 · OpenAI · observed 2026-08-27 | 10,000 five-minute support sessions; fixed task-success floor; accepted resolutions | 50,000 minutes; audio I/O $0.012/min blended; model/tools $0.019/session; success 92.1% ≥ 90% QUALIFIED — 50,000×$0.012+10,000×$0.019 = $790.00; $0.079/resolution |
| batch25-cheap-m3-r3 · xAI · observed 2026-08-27 | 10,000 five-minute support sessions; fixed task-success floor; accepted resolutions | Connection/audio/reconnect rates present; task-success 88.6% below the 90% floor $742.00 observed; EXCLUDED — quality floor failed |
| batch25-cheap-m3-r4 · Anthropic · observed 2026-08-27 | 10,000 five-minute support sessions; fixed task-success floor; accepted resolutions | Computer-use screenshot rates do not provide realtime audio I/O units EXCLUDED — incompatible modality |
| batch25-cheap-m3-r5 · DeepSeek · observed 2026-08-27 | 10,000 five-minute support sessions; fixed task-success floor; accepted resolutions | No connection, audio I/O, or interruption rates in dated registry EXCLUDED — missing specialized evidence |
| batch25-cheap-m3-r6 · Amazon · observed 2026-08-27 | 10,000 five-minute support sessions; fixed task-success floor; accepted resolutions | No complete connection/audio/transcription/tool rate tuple in frozen record EXCLUDED — incomplete unit set |
Formula / qualification rule: rank only candidates with every required unit, rounding/retry rule, and matched quality or success floor. Ordering for this frozen run is Google > OpenAI; all excluded candidates remain below the qualified set. Specialized rate source: dated evidence index 2026-08-27; base text, embeddings, transcription, and generic audio rates are not substituted.
Verified 2026-08-08. Data owner: Luna. Run the cheapest-ai-api evidence scenario →
Batch 26 · moderation, video-understanding, and speech-translation ladders
Frozen verification window: 2026-08-27 UTC. Candidates are independently excluded when a specialized unit, rate, or matched quality floor is absent.
1. Content-moderation cheapest-qualified ladder
Frozen workload: 10M multilingual text/image items; fixed false-allow/false-block floor; accepted decisions
| Candidate / run | Field observation | Decision | Cost / exclusion |
|---|---|---|---|
OpenAIbatch26-cheapest-m1-r1observed 2026-08-27 | text endpoint eligible; image item limit 1,000; matched quality 98.1% | QUALIFIED — 9,992,000 accepted decisions | $0.00 ÷ 9,992,000 = $0.000000/decision |
Googlebatch26-cheapest-m1-r2observed 2026-08-27 | text eligible; image endpoint rate absent for the frozen queue | EXCLUDED — incompatible dated image rate | Unavailable — Google image moderation rate |
Anthropicbatch26-cheapest-m1-r3observed 2026-08-27 | no moderation endpoint/item unit in dated registry | EXCLUDED — incompatible endpoint | Unavailable — Anthropic moderation endpoint rate |
DeepSeekbatch26-cheapest-m1-r4observed 2026-08-27 | no matched moderation quality run | EXCLUDED — missing matched floor evidence | Unavailable — DeepSeek moderation quality run |
xAIbatch26-cheapest-m1-r5observed 2026-08-27 | text moderation evidence only; image item tuple absent | EXCLUDED — incomplete modality tuple | Unavailable — xAI image moderation unit |
Amazonbatch26-cheapest-m1-r6observed 2026-08-27 | request/item rate and matched false-allow review absent | EXCLUDED — missing specialized record | Unavailable — Amazon moderation rate and matched run |
Formula / qualification rule: Rank only candidates with compatible request/item/image units and matched moderation quality; cost per accepted decision = total attributable bill ÷ accepted decisions. Source: pricing registry and dated evidence index verified 2026-08-27.
2. Video-understanding cheapest-qualified ladder
Frozen workload: 10,000 hours; declared resolution/sampling; timestamp-localization floor; accepted clips
| Candidate / run | Field observation | Decision | Cost / exclusion |
|---|---|---|---|
Googlebatch26-cheapest-m2-r1observed 2026-08-27 | video/audio units and 91.4% timestamp accuracy matched; 9,812 hours accepted | QUALIFIED — complete unit tuple | $0.004/min video + sourced audio/tool units; exact total $1,842.16 |
OpenAIbatch26-cheapest-m2-r2observed 2026-08-27 | image-input rate exists; no video unit or timestamp run | EXCLUDED — image rate cannot substitute for video | Unavailable — OpenAI video unit and matched localization run |
Anthropicbatch26-cheapest-m2-r3observed 2026-08-27 | image/document evidence; no video rate | EXCLUDED — incompatible modality | Unavailable — Anthropic video rate |
DeepSeekbatch26-cheapest-m2-r4observed 2026-08-27 | no video endpoint/unit in dated registry | EXCLUDED — missing specialized rate | Unavailable — DeepSeek video rate |
xAIbatch26-cheapest-m2-r5observed 2026-08-27 | image/video support claim without matched 10,000-hour bill | EXCLUDED — missing matched invoice evidence | Unavailable — xAI video invoice and localization run |
Amazonbatch26-cheapest-m2-r6observed 2026-08-27 | video input rate present but no timestamp-localization floor | EXCLUDED — missing matched quality evidence | Unavailable — Amazon timestamp-localization run |
Formula / qualification rule: Cost per accepted hour = video + audio + text + upload/storage/tool/retry units ÷ accepted hours; image-only rates cannot qualify. Source: pricing registry and dated evidence index verified 2026-08-27.
3. Speech-translation cheapest-qualified ladder
Frozen workload: 1,000 multilingual hours; fixed WER/semantic-adequacy floors; accepted translated hours
| Candidate / run | Field observation | Decision | Cost / exclusion |
|---|---|---|---|
Googlebatch26-cheapest-m3-r1observed 2026-08-27 | transcription/translation/audio units sourced; WER 6.2%; adequacy 94.1%; 984 hours accepted | QUALIFIED — complete tuple and quality floor | $1,264.32 ÷ 984 = $1.2849/accepted hour |
OpenAIbatch26-cheapest-m3-r2observed 2026-08-27 | transcription and text rates; no matched multilingual translation adequacy run | EXCLUDED — missing specialized quality evidence | Unavailable — OpenAI speech-translation matched run |
Anthropicbatch26-cheapest-m3-r3observed 2026-08-27 | text generation rate; no transcription/audio-duration unit | EXCLUDED — text rate cannot substitute for speech | Unavailable — Anthropic transcription and audio units |
DeepSeekbatch26-cheapest-m3-r4observed 2026-08-27 | translation text rate only; no transcription unit | EXCLUDED — incomplete modality tuple | Unavailable — DeepSeek transcription rate |
xAIbatch26-cheapest-m3-r5observed 2026-08-27 | audio input present; translation output/adequacy run absent | EXCLUDED — missing output and quality evidence | Unavailable — xAI speech-translation output rate and run |
Amazonbatch26-cheapest-m3-r6observed 2026-08-27 | transcription rate present; no matched semantic-adequacy floor | EXCLUDED — transcription-only evidence | Unavailable — Amazon translation quality run |
Formula / qualification rule: Cost per accepted hour = transcription + translation + audio-duration/rounding + diarization + retry/output units ÷ accepted hours; TTS-only rates cannot qualify. Source: pricing registry and dated evidence index verified 2026-08-27.
Verified 2026-08-08. Data owner: Luna. Run the cheapest-ai-api Batch 26 evidence scenario →
Batch 27 · document-translation, OCR/layout, and video-generation ladders
Frozen verification window: 2026-08-27 UTC. Candidates are independently excluded when a specialized unit, artifact-quality floor, or matched acceptance run is absent.
1. Document-translation cheapest-qualified ladder
Frozen workload: 10M words; 12 languages; DOCX, HTML, PDF; layout/glossary and adequacy floor
| Candidate / run | Visible inputs | Field-level observation | Decision boundary | Cost / exclusion |
|---|---|---|---|---|
Googlebatch27-cheapest-m1-r1observed 2026-08-27 | document + translation units; 12 languages; adequacy 94.2% | 9.86M accepted words; DOCX layout 98.1%; 1.2% human review | QUALIFIED — complete compatible tuple | $18,422.60 ÷ 9.86 = $1,868.42/accepted M words |
OpenAIbatch27-cheapest-m1-r2observed 2026-08-27 | text tokens present; no matched document-layout unit | semantic output available; PDF layout acceptance absent | EXCLUDED — base text rate cannot substitute | Unavailable — document unit and layout/adequacy run |
Anthropicbatch27-cheapest-m1-r3observed 2026-08-27 | text output; no sourced document-translation unit | DOCX/PDF packet not costed on compatible unit | EXCLUDED — missing specialized rate | Unavailable — document-translation unit and matched run |
Formula / qualification rule: Cost per accepted million words = sourced input/output/document/translation units + retries + review ÷ accepted words; base text and speech rates cannot qualify. Source: pricing registry and dated evidence index verified 2026-08-27.
2. OCR-and-layout-reconstruction cheapest-qualified ladder
Frozen workload: 1M pages; clean text, tables, forms, handwriting, rotated scans; field/reading-order floors
| Candidate / run | Visible inputs | Field-level observation | Decision boundary | Cost / exclusion |
|---|---|---|---|---|
Googlebatch27-cheapest-m2-r1observed 2026-08-27 | page/image units; batch; table and field error floors | 982,400 accepted pages; table error 2.1%; rotated scans 96.4%; $0.004/page equivalent | QUALIFIED — compatible OCR and layout evidence | $3,929.60 ÷ 982,400 = $0.004000/page |
OpenAIbatch27-cheapest-m2-r2observed 2026-08-27 | image input rate; no OCR field/reading-order run | vision answers present; handwriting and table floor absent | EXCLUDED — generic vision cannot substitute | Unavailable — OCR-specific unit and matched layout run |
xAIbatch27-cheapest-m2-r3observed 2026-08-27 | image input; upload/asset rate missing | clean text sample only; no 1M-page accepted denominator | EXCLUDED — incomplete compatible tuple | Unavailable — OCR asset/storage units and matched run |
Formula / qualification rule: Cost per accepted page = page/image/token/tool/upload/storage/retry units ÷ accepted pages; generic vision/extraction rates never substitute. Source: pricing registry and dated evidence index verified 2026-08-27.
3. Video-generation cheapest-qualified ladder
Frozen workload: 10,000 accepted 5-second 1080p clips; audio, safety, retry, retention, temporal/prompt floors
| Candidate / run | Visible inputs | Field-level observation | Decision boundary | Cost / exclusion |
|---|---|---|---|---|
Runwaybatch27-cheapest-m3-r1observed 2026-08-27 | 1080p/5s; audio; safety rejection; temporal score; retention | 9,812 clips accepted; 49,060 seconds; prompt adherence 92.1%; 188 rejected | QUALIFIED — full generation tuple and acceptance floor | $12,265.00 ÷ 49,060 = $0.250000/accepted second |
OpenAIbatch27-cheapest-m3-r2observed 2026-08-27 | image generation rate; no matched 1080p video-generation tuple | image artifacts only; no accepted video denominator | EXCLUDED — image-generation rate cannot substitute | Unavailable — video-generation duration/resolution rate and run |
Googlebatch27-cheapest-m3-r3observed 2026-08-27 | video understanding units; generation output rate absent | analysis endpoint available; generated artifact retention absent | EXCLUDED — video-understanding rate cannot qualify | Unavailable — video-generation output rate and artifact run |
Formula / qualification rule: Cost per accepted second = duration/resolution/credit/token/tool + failed/cancelled/retry/retention units ÷ accepted seconds; image/video-understanding rates are barred. Source: pricing registry and dated evidence index verified 2026-08-27.
Verified 2026-08-08. Data owner: Luna. Run the cheapest-ai-api Batch 27 evidence scenario →
Batch 28 · Specialist cheapest-qualified API ladders
Frozen verification window: 2026-08-27 UTC. These are server-rendered matched fixtures, not live estimates. Each row exposes inputs, a reproducible formula/result or a narrowly scoped missing record, dated provenance, and a decision boundary.
1. Source-code vulnerability-triage cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level observation | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
Candidate Abatch28-cheapest-m1-r1observed 2026-08-27 | 100,000 findings; multilingual; CWE/severity/location floor | 96,400 accepted; severity 94.1%; false-dismissal 1.8%; full unit tuple | QUALIFIED — $8,676.00 ÷ 96,400 accepted findings | $0.090000/accepted finding |
Candidate Bbatch28-cheapest-m1-r2observed 2026-08-27 | token rate present; parser/sandbox unit absent | quality sample passes but compatible denominator incomplete | EXCLUDED — missing specialized unit | Unavailable — parser/sandbox rate and matched finding run |
Candidate Cbatch28-cheapest-m1-r3observed 2026-08-27 | generic coding benchmark only | no vulnerability false-dismissal floor | EXCLUDED — coding score cannot qualify triage | Unavailable — vulnerability-specific quality and cost tuple |
Formula / scoring rule: Cost per accepted finding = sourced token/tool/cache/Batch/parser/sandbox/retry units ÷ accepted findings; generic coding and moderation rates cannot qualify. Source: pricing registry and dated evidence index verified 2026-08-27.
2. Meeting diarization-and-action-extraction cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level observation | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
Candidate Abatch28-cheapest-m2-r1observed 2026-08-27 | 10,000 hours; overlap; speaker and action/date floors | 9,720 accepted hours; speaker attribution 93.8%; action owner/date 91.2% | QUALIFIED — $29,160.00 ÷ 9,720 hours | $3.000000/accepted hour |
Candidate Bbatch28-cheapest-m2-r2observed 2026-08-27 | transcription and token rates; diarization unit absent | WER measured; speaker attribution not measured | EXCLUDED — transcription-only result cannot substitute | Unavailable — diarization unit and matched speaker run |
Candidate Cbatch28-cheapest-m2-r3observed 2026-08-27 | speech-translation result; action extraction absent | translation adequacy passes; action-owner floor missing | EXCLUDED — translation cannot qualify extraction | Unavailable — action-extraction quality and cost tuple |
Formula / scoring rule: Cost per accepted hour = audio duration + transcription/diarization/model/storage/retry units ÷ accepted hours under speaker/action accuracy floors. Source: pricing registry and dated evidence index verified 2026-08-27.
3. Document PII-redaction cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level observation | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
Candidate Abatch28-cheapest-m3-r1observed 2026-08-27 | 1,000,000 digital/scanned pages; entity and layout floors | 982,000 accepted; entity recall 98.4%; over-redaction 1.1%; layout 97.8% | QUALIFIED — $49,100.00 ÷ 982,000 pages | $0.050000/accepted page |
Candidate Bbatch28-cheapest-m3-r2observed 2026-08-27 | OCR/page rate; redaction span review absent | OCR quality reported; PII recall and over-redaction not run | EXCLUDED — OCR ladder cannot supply redaction verdict | Unavailable — PII entity-span run and compatible redaction units |
Candidate Cbatch28-cheapest-m3-r3observed 2026-08-27 | writing privacy rewrite evidence | text rewrite accepted; scanned layout not tested | EXCLUDED — writing privacy is not document redaction | Unavailable — page/image/layout redaction run |
Formula / scoring rule: Cost per accepted page = page/image/OCR/token/tool/storage/retry units ÷ accepted pages under entity recall, over-redaction, and layout floors. Source: pricing registry and dated evidence index verified 2026-08-27.
Verified 2026-08-08. Data owner: Luna. Missing specialized units, rates, and matched runs are never inferred from a neighboring modality or provider. Run the cheapest Batch 28 evidence scenario →
Batch 29 · Specialist cheapest-qualified API ladders
Frozen verification window: 2026-08-27 UTC. These are server-rendered matched fixtures, not live estimates. Each row exposes frozen inputs, a reproducible formula/result or a narrowly scoped missing record, dated provenance, and a decision boundary.
1. Clinical-code suggestion cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level observation | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
Candidate Abatch29-cheapest-m1-r1observed 2026-08-27 | 100,000 de-identified ICD-style notes; expert-review floor; full unit tuple | 96,400 accepted; exact-code and hierarchy floors pass; unsupported/missed ceilings pass | QUALIFIED — workflow evidence does not make a care decision | $12,050.00 ÷ 96,400 = $0.125000/accepted suggestion |
Candidate Bbatch29-cheapest-m1-r2observed 2026-08-27 | token/cache/Batch rates; terminology or expert-review unit absent | quality sample exists but compatible accepted denominator is incomplete | EXCLUDED — missing specialized unit and review cost | Unavailable — terminology/retrieval and mandatory expert-review cost tuple |
Candidate Cbatch29-cheapest-m1-r3observed 2026-08-27 | generic extraction benchmark; no exact-code hierarchy gate | extraction quality cannot establish clinical-code qualification | EXCLUDED — generic extraction cannot qualify the ladder | Unavailable — exact-code quality, unsupported-code ceiling, and matched bill |
Formula / scoring rule: Cost per expert-accepted suggestion = sourced token/cache/Batch/tool/terminology/retrieval/retry cost ÷ accepted suggestions; exact-code, hierarchy, unsupported-code, and missed-code gates must pass. Source: pricing registry and dated evidence index verified 2026-08-27.
2. Legal e-discovery privilege-review cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level observation | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
Candidate Abatch29-cheapest-m2-r1observed 2026-08-27 | 1,000,000 emails/attachments/OCR pages; threading/dedupe; counsel sample | 962,000 accepted; recall and false-withhold floors pass; rationale spans trace | QUALIFIED — counsel review remains mandatory | $57,720.00 ÷ 962,000 = $0.060000/accepted document |
Candidate Bbatch29-cheapest-m2-r2observed 2026-08-27 | OCR/page/token rates; privilege-review sample absent | OCR and dedupe pass but privilege recall cannot be qualified | EXCLUDED — PII or summary evidence cannot substitute | Unavailable — privilege recall, false-withhold, counsel sample, and compatible units |
Candidate Cbatch29-cheapest-m2-r3observed 2026-08-27 | document summary output; rationale-to-source span absent | summary quality does not establish privilege evidence | EXCLUDED — no accepted-document denominator | Unavailable — privilege rationale spans and counsel-accepted document run |
Formula / scoring rule: Cost per accepted document = document/image/OCR/token/search/storage/retry cost ÷ counsel-accepted documents after privilege recall and false-withhold gates; no legal decision is made. Source: pricing registry and dated evidence index verified 2026-08-27.
3. Insurance property-damage triage cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level observation | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
Candidate Abatch29-cheapest-m3-r1observed 2026-08-27 | 250,000 claim packets; photos/notes/invoices/policy excerpts; adjuster floor | 238,500 accepted; damage/severity and evidence-span floors pass; duplicates controlled | QUALIFIED — triage evidence is not coverage or payment advice | $71,550.00 ÷ 238,500 = $0.300000/accepted packet |
Candidate Bbatch29-cheapest-m3-r2observed 2026-08-27 | vision/OCR/token rates; adjuster acceptance and coverage ceiling absent | damage labels reported but unsupported-coverage ceiling is unmeasured | EXCLUDED — generic vision cannot qualify triage | Unavailable — adjuster acceptance, coverage-claim ceiling, and compatible unit tuple |
Candidate Cbatch29-cheapest-m3-r3observed 2026-08-27 | writing/extraction result; photos and duplicate-claim gate absent | narrative quality cannot establish packet qualification | EXCLUDED — no accepted packet denominator | Unavailable — image/document evidence, duplicate detection, and adjuster run |
Formula / scoring rule: Cost per accepted packet = sourced image/document/OCR/token/tool/storage/retry cost ÷ adjuster-accepted packets after damage/severity, evidence-span, duplicate, and unsupported-coverage gates; no payment decision is made. Source: pricing registry and dated evidence index verified 2026-08-27.
Verified 2026-08-08. Data owner: Luna. Missing specialized units, rates, and matched runs are never inferred from a neighboring modality, provider, or prior batch. Run the cheapest Batch 29 evidence scenario →
Batch 30 · Specialist cheapest-qualified API ladders
Frozen verification window: 2026-08-27 UTC. These server-rendered fixtures expose inputs, formulas, field-level observations, decision boundaries, dated provenance, and exact bills where the registry closes the token tuple. Missing specialist evidence is explicitly Unavailable.
1. Customs-and-shipping-document validation cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level observation | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
1 · DeepSeek-V3 — qualifiedbatch30-cheapest-m1-r1observed 2026-08-27 | batch30-cheapest-m1-r1; 1,000 packets; 4 pages/packet; 2026-08-27T08:00Z | OCR 4,000 pages × $0.0015=$6.000000; 1,240,000 input × $0.28/1M=$0.347200; 180,000 output × $0.42/1M=$0.075600; schema/tool $0.220000; retries $0.140000; total $6.782800; 962 broker-accepted packets; 95.2% identifier agreement, 1.8% missing-field, 0.7% unsupported-classification. | QUALIFIED #1 — $0.007050/accepted packet; cheapest observed provider | $6.782800 ÷ 962 = $0.007050 per accepted packet; unit order: OCR → tokens → schema/tool → retry |
2 · OpenAI GPT-4o-mini — qualifiedbatch30-cheapest-m1-r2observed 2026-08-27 | batch30-cheapest-m1-r2; same corpus; 2026-08-27T08:18Z | OCR 4,000 pages × $0.002=$8.000000; 1,180,000 input × $0.50/1M=$0.590000; 165,000 output × $1.50/1M=$0.247500; schema/tool $0.310000; retries $0.180000; total $9.327500; 970 accepted; agreement 96.0%, missing 1.4%, unsupported 0.6%. | QUALIFIED #2 — $0.009616/accepted packet | $9.327500 ÷ 970 = $0.009616 per accepted packet; unit order: OCR → tokens → schema/tool → retry |
3 · Google Gemini 2.0 Flash — excludedbatch30-cheapest-m1-r3observed 2026-08-27 | batch30-cheapest-m1-r3; same corpus; 2026-08-27T08:36Z | Token bill $4.912400 and OCR bill $5.600000 are returned, but 91 packets lack broker review and unsupported classification is 2.4%. | EXCLUDED — acceptance ceiling breached | $10.512400 measured workflow spend; no qualified denominator |
Formula / scoring rule: Cost per accepted packet = sourced page/image/OCR/token/schema/tool/storage/retry units ÷ broker-accepted packets after cross-document agreement, missing-field, and unsupported-classification ceilings. No tariff or admissibility decision is made. Source: pricing registry and dated evidence index verified 2026-08-27; provider registry: DeepSeek-V3 / OCR and token registry record verified 2026-08-27OpenAI GPT-4o-mini / OCR and token registry record verified 2026-08-27Google Gemini 2.0 Flash / OCR and token registry record verified 2026-08-27; test suite: Batch 30 customs/shipping validation cheapest-qualified test suite (run and result recorded 2026-08-27).
2. Contact-center compliance-QA cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level observation | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
1 · OpenAI GPT-4o-mini — qualifiedbatch30-cheapest-m2-r1observed 2026-08-27 | batch30-cheapest-m2-r1; 10,000 one-hour calls; 2026-08-27T08:55Z | Transcription 10,000 h × $0.006=$60.000000; diarization $8.000000; 2,800,000 input × $0.50/1M=$1.400000; 420,000 output × $1.50/1M=$0.630000; retrieval/storage $4.200000; retry $1.100000; total $75.330000; 9,420 supervisor-accepted; disclosure recall 98.1%, attribution 97.4%, false flags 1.2%. | QUALIFIED #1 — $0.007997/accepted interaction | $75.330000 ÷ 9420 = $0.007997; units: audio → diarization → tokens → retrieval/storage → retry |
2 · DeepSeek-V3 — qualifiedbatch30-cheapest-m2-r2observed 2026-08-27 | batch30-cheapest-m2-r2; same corpus; 2026-08-27T09:14Z | Audio $70.000000; diarization $10.000000; 2,460,000 input × $0.28/1M=$0.688800; 390,000 output × $0.42/1M=$0.163800; retrieval/storage $3.800000; retry $0.900000; total $85.552600; 9,180 accepted; recall 97.8%, attribution 96.9%, false flags 1.6%. | QUALIFIED #2 — $0.009320/accepted interaction | $85.552600 ÷ 9180 = $0.009320; all three QA ceilings pass |
3 · Google Gemini 2.0 Flash — excludedbatch30-cheapest-m2-r3observed 2026-08-27 | batch30-cheapest-m2-r3; multilingual fallback; 2026-08-27T09:33Z | Measured total $68.414000 but 8,740/10,000 supervisor rows accepted; false-flag rate 3.9% exceeds 2.0% ceiling. | EXCLUDED — compliance false-flag ceiling breached | $68.414000 measured; not divided into a qualified result |
Formula / scoring rule: Cost per accepted interaction = sourced audio-duration/transcription/diarization/token/retrieval/storage/retry units ÷ supervisor-accepted interactions after disclosure/consent recall, attribution, and false-flag ceiling. Source: pricing registry and dated evidence index verified 2026-08-27; provider registry: OpenAI GPT-4o-mini / audio and token registry record verified 2026-08-27DeepSeek-V3 / audio and token registry record verified 2026-08-27Google Gemini 2.0 Flash / audio and token registry record verified 2026-08-27; test suite: Batch 30 contact-center compliance-QA cheapest-qualified test suite (run and result recorded 2026-08-27).
3. Satellite-imagery change-triage cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level observation | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
1 · DeepSeek-V3 — qualifiedbatch30-cheapest-m3-r1observed 2026-08-27 | batch30-cheapest-m3-r1; 5,000 before/after 1024px pairs; 2026-08-27T09:52Z | 10,000 tiles × $0.0008=$8.000000; 1,840,000 input × $0.28/1M=$0.515200; 260,000 output × $0.42/1M=$0.109200; tool $1.500000; storage/retry $0.940000; total $11.064400; 4,620 analyst-accepted pairs; registration 0.86px, change recall 94.2%, false-change 1.7%. | QUALIFIED #1 — $0.002395/accepted pair | $11.064400 ÷ 4620 = $0.002395; units: tiles → tokens → tool → storage/retry |
2 · OpenAI GPT-4o-mini — qualifiedbatch30-cheapest-m3-r2observed 2026-08-27 | batch30-cheapest-m3-r2; same pairs; 2026-08-27T10:10Z | 10,000 tiles × $0.001=$10.000000; 2,120,000 input × $0.50/1M=$1.060000; 300,000 output × $1.50/1M=$0.450000; tool $1.800000; storage/retry $1.200000; total $14.510000; 4,580 accepted; registration 0.91px, recall 95.0%, false-change 1.5%. | QUALIFIED #2 — $0.003168/accepted pair | $14.510000 ÷ 4580 = $0.003168; all spatial and reviewer gates pass |
3 · Google Gemini 2.0 Flash — excludedbatch30-cheapest-m3-r3observed 2026-08-27 | batch30-cheapest-m3-r3; same pairs; 2026-08-27T10:28Z | Measured total $9.882000; only 4,210 analyst acceptances; registration p95 1.8px exceeds 1.5px tolerance and localization missing on 6.4%. | EXCLUDED — registration/localization gates breached | $9.882000 measured; no qualified pair cost |
Formula / scoring rule: Cost per accepted pair = sourced image/tile/token/tool/storage/retry units ÷ analyst-accepted pairs after registration tolerance, change-class recall, false-change ceiling, and evidence localization. Source: pricing registry and dated evidence index verified 2026-08-27; provider registry: DeepSeek-V3 / tile and token registry record verified 2026-08-27OpenAI GPT-4o-mini / tile and token registry record verified 2026-08-27Google Gemini 2.0 Flash / tile and token registry record verified 2026-08-27; test suite: Batch 30 satellite change-triage cheapest-qualified test suite (run and result recorded 2026-08-27).
Verified 2026-08-08. Data owner: Luna. Prior-batch, adjacent-suite, provider, and unsupported fields are not substituted. Run the cheapest Batch 30 evidence scenario →
Batch 31 · Specialist cheapest-qualified API ladders
Frozen verification window: 2026-08-27 UTC. Matched model/run identity, inputs, formulas, field-level observations, decision boundaries, dated provenance, and exact bills are server-rendered. Unsupported fields fail closed as Unavailable.
1. Pharmacovigilance adverse-event case-intake cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level observation | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
1 · DeepSeek-V3 — qualifiedbatch31-cheapest-m1-r1model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 1,000,000 reports; 4 pages/report; candidate named; 07:04Z | OCR $1,500.000000 + tokens $420.000000 + schema/retrieval/storage/retry $180.000000 = $2,100.000000; 932,000 accepted; field recall 96.2%, duplicate linkage 94.1%. | QUALIFIED #1 — $0.002253 per accepted intake packet | $2,100.000000 ÷ 932000 = $0.002253; accepted denominator is reviewer packets, not reports |
2 · OpenAI GPT-4o-mini — qualifiedbatch31-cheapest-m1-r2model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same 1,000,000-report corpus; candidate named; 07:22Z | OCR $2,000.000000 + tokens $1,140.000000 + schema/retrieval/storage/retry $260.000000 = $3,400.000000; 956,000 accepted; recall 97.4%, linkage 95.0%. | QUALIFIED #2 — $0.003556 per accepted intake packet | $3,400.000000 ÷ 956000 = $0.003556; human acceptance denominator exposed |
3 · Google Gemini 2.0 Flash — excludedbatch31-cheapest-m1-r3model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; candidate named; 07:40Z | Measured OCR/tokens/workflow total $1,880.000000; 884,000 accepted; source-span recall 91.2% below 94.0% floor. | EXCLUDED — source-span gate breached; no qualified cost | $1,880.000000 measured; accepted denominator is 884000 but ladder cost is Unavailable because gate failed |
Formula / scoring rule: Cost per accepted intake packet = sourced OCR/audio/page/token/schema/retrieval/storage/retry units ÷ safety-reviewer-accepted packets, after field-recall, duplicate-linkage, and source-span gates. No causality or seriousness decision is made. First-party registry: allaiask.com pricing and evidence registry, verified 2026-08-27. Provider/model source: Named candidates: DeepSeek-V3, OpenAI GPT-4o-mini, Google Gemini 2.0 Flash; pharmacovigilance pricing registry, verified 2026-08-27.
2. Software bill-of-materials and license-obligation extraction cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level observation | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
1 · DeepSeek-V3 — qualifiedbatch31-cheapest-m2-r1model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 100,000 repositories; manifests/lockfiles/binaries; 08:02Z | Files $640.000000 + tokens $210.000000 + search/tool/storage/retry $150.000000 = $1,000.000000; 91,200 accepted; hash recall 97.1%, edge precision 96.4%. | QUALIFIED #1 — $0.010965 per accepted repository | $1,000.000000 ÷ 91200 = $0.010965; 2,400 ambiguous licenses escalated |
2 · OpenAI GPT-4o-mini — qualifiedbatch31-cheapest-m2-r2model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; candidate named; 08:20Z | Files $820.000000 + tokens $480.000000 + search/tool/storage/retry $220.000000 = $1,520.000000; 93,600 accepted; hash recall 98.0%, edge precision 97.2%. | QUALIFIED #2 — $0.016239 per accepted repository | $1,520.000000 ÷ 93600 = $0.016239; license citations retained for review |
3 · Google Gemini 2.0 Flash — excludedbatch31-cheapest-m2-r3model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; candidate named; 08:38Z | Measured total $760.000000; 89,100 accepted; binary-evidence recall 89.8% below 93.0% floor; 6,100 escalations. | EXCLUDED — binary evidence gate breached | $760.000000 measured; accepted repository cost Unavailable because qualification failed |
Formula / scoring rule: Cost per accepted repository = sourced file/token/code-search/tool/storage/retry units ÷ reviewer-accepted repositories after package/hash recall, dependency precision, license citation, and escalation gates. No legal advice is given. First-party registry: allaiask.com pricing and evidence registry, verified 2026-08-27. Provider/model source: Named candidates: DeepSeek-V3, OpenAI GPT-4o-mini, Google Gemini 2.0 Flash; SBOM pricing registry, verified 2026-08-27.
3. Patent prior-art landscape screening cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level observation | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
1 · DeepSeek-V3 — qualifiedbatch31-cheapest-m3-r1model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 10,000,000 multilingual documents; claims/drawings/families; 08:56Z | OCR/image $4,200.000000 + translation/tokens $2,100.000000 + search/vector/storage/retry $1,200.000000 = $7,500.000000; 7,820 analyst-accepted sets. | QUALIFIED #1 — $0.959079 per accepted candidate set | $7,500.000000 ÷ 7820 = $0.959079; family dedup 98.1%, claim evidence recall 94.0% |
2 · OpenAI GPT-4o-mini — qualifiedbatch31-cheapest-m3-r2model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same 10,000,000-document corpus; candidate named; 09:14Z | OCR/image $5,100.000000 + translation/tokens $3,400.000000 + search/vector/storage/retry $1,800.000000 = $10,300.000000; 8,140 accepted sets. | QUALIFIED #2 — $1.265356 per accepted candidate set | $10,300.000000 ÷ 8140 = $1.265356; date/jurisdiction fidelity 97.2% |
3 · Google Gemini 2.0 Flash — excludedbatch31-cheapest-m3-r3model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; candidate named; 09:32Z | Measured total $6,800.000000; 7,100 accepted sets; multilingual claim-span precision 88.9% below 92.0% floor. | EXCLUDED — claim-span precision gate breached | $6,800.000000 measured; accepted-set cost Unavailable because qualification failed |
Formula / scoring rule: Cost per accepted candidate set = sourced OCR/image/translation/token/search/vector/storage/retry units ÷ analyst-accepted candidate sets after family/date/jurisdiction/claim-evidence gates. No novelty, validity, FTO, or infringement determination is made. First-party registry: allaiask.com pricing and evidence registry, verified 2026-08-27. Provider/model source: Named candidates: DeepSeek-V3, OpenAI GPT-4o-mini, Google Gemini 2.0 Flash; patent-landscape pricing registry, verified 2026-08-27.
Verified 2026-08-08. Data owner: Luna. Prior-batch, adjacent-suite, provider, and unsupported fields are not substituted. Run the cheapest Batch 31 evidence scenario →
Batch 32 · Catalog, planning-permit, and trade-surveillance cheapest-qualified ladders
Frozen verification window: 2026-08-27 UTC. Matched model/run identity, frozen inputs, formulas, field-level observations, decision boundaries, dated provenance, and exact bills are server-rendered. Unsupported fields fail closed as Unavailable.
1. Product-catalog normalization cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level observation | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
1 · DeepSeek-V3 / cat32-1011batch32-cheapest-m1-r1model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 10M SKUs; images, supplier sheets, taxonomy; candidate named; 07:14Z | DeepSeek-V3: image/page/OCR/token/schema/embedding/search/storage/retry = $8,420.000000; 9,420,000 accepted; identifier recall 97.1%. | QUALIFIED #1 — $0.000894 per accepted SKU | $8,420.000000 ÷ 9420000 = $0.000894; human merchandiser denominator exposed |
2 · OpenAI GPT-4o-mini / cat32-1012batch32-cheapest-m1-r2model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same 10M-SKU feed; candidate named; 07:30Z | OpenAI GPT-4o-mini: compatible sourced units total $12,680.000000; 9,610,000 accepted; taxonomy precision 96.4%. | QUALIFIED #2 — $0.001320 per accepted SKU | $12,680.000000 ÷ 9610000 = $0.001320; accepted SKU denominator is reviewer decisions |
3 · Google Gemini 2.0 Flash / cat32-1013batch32-cheapest-m1-r3model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same feed; candidate named; 07:46Z | Google Gemini 2.0 Flash: sourced units total $7,940.000000; 8,610,000 accepted; duplicate-cluster quality 89.1% below 93% floor. | EXCLUDED — duplicate-quality gate breached; no qualified cost | Unavailable — Google Gemini 2.0 Flash qualified cost is unavailable after gate failure; 8610000 accepted is retained |
Formula / scoring rule: Cost per accepted SKU = sourced image/page/OCR/token/schema/embedding/search/storage/retry units ÷ merchandiser-accepted SKUs after recall/precision/duplicate gates. First-party pricing/evidence registry: DeepSeek-V3, OpenAI GPT-4o-mini, Google Gemini 2.0 Flash product-catalog pricing registry, verified 2026-08-27; unsupported usage or credits remain Unavailable.. Dated registry and evidence index, verified 2026-08-27.
2. Planning-permit packet completeness cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level observation | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
1 · DeepSeek-V3 / per32-1021batch32-cheapest-m2-r1model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 500K applications; forms/drawings/maps/revisions; candidate named; 08:02Z | DeepSeek-V3: compatible page/image/OCR/token/retrieval/vector/storage/retry units = $6,240.000000; 462,000 accepted; required-item recall 96.0%. | QUALIFIED #1 — $0.013506 per accepted packet | $6,240.000000 ÷ 462000 = $0.013506; planner-review denominator exposed |
2 · OpenAI GPT-4o-mini / per32-1022batch32-cheapest-m2-r2model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same 500K corpus; candidate named; 08:18Z | OpenAI GPT-4o-mini: compatible sourced units = $9,860.000000; 474,000 accepted; parcel/date fidelity 97.2%. | QUALIFIED #2 — $0.020802 per accepted packet | $9,860.000000 ÷ 474000 = $0.020802; human acceptance is not application count |
3 · Google Gemini 2.0 Flash / per32-1023batch32-cheapest-m2-r3model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; candidate named; 08:34Z | Google Gemini 2.0 Flash: sourced units = $5,880.000000; 421,000 accepted; evidence-span precision 88.7% below 92% floor. | EXCLUDED — evidence-span gate breached; no qualified cost | Unavailable — Google Gemini 2.0 Flash qualified packet cost is unavailable after gate failure; 421000 accepted is retained |
Formula / scoring rule: Cost per accepted packet = sourced page/image/OCR/token/retrieval/vector/storage/retry units ÷ planner-accepted packets after version, parcel/date, and evidence-span gates. First-party pricing/evidence registry: DeepSeek-V3, OpenAI GPT-4o-mini, Google Gemini 2.0 Flash planning-permit pricing registry, verified 2026-08-27; unsupported usage or credits remain Unavailable.. Dated registry and evidence index, verified 2026-08-27.
3. Trade-surveillance case-assembly cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level observation | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
1 · DeepSeek-V3 / trd32-1031batch32-cheapest-m3-r1model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 1B events; orders/executions/chats/calls/notices; candidate named; 08:50Z | DeepSeek-V3: compatible audio/transcription/token/search/vector/storage/tool/retry units = $48,200.000000; 812,000 accepted; time-sequence linkage 95.4%. | QUALIFIED #1 — $0.059360 per accepted case packet | $48,200.000000 ÷ 812000 = $0.059360; investigator denominator exposed |
2 · OpenAI GPT-4o-mini / trd32-1032batch32-cheapest-m3-r2model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same 1B-event corpus; candidate named; 09:06Z | OpenAI GPT-4o-mini: compatible sourced units = $72,400.000000; 846,000 accepted; citation-span traceability 96.1%. | QUALIFIED #2 — $0.085579 per accepted case packet | $72,400.000000 ÷ 846000 = $0.085579; accepted cases are human-reviewed packets |
3 · Google Gemini 2.0 Flash / trd32-1033batch32-cheapest-m3-r3model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; candidate named; 09:22Z | Google Gemini 2.0 Flash: sourced units = $41,600.000000; 704,000 accepted; false-association ceiling breached at 8.4% vs 5% floor. | EXCLUDED — false-association gate breached; no misconduct conclusion | Unavailable — Google Gemini 2.0 Flash qualified case cost is unavailable after gate failure; 704000 accepted is retained |
Formula / scoring rule: Cost per accepted case packet = sourced audio/transcription/token/search/vector/storage/tool/retry units ÷ investigator-accepted packets after entity/time/evidence gates. First-party pricing/evidence registry: DeepSeek-V3, OpenAI GPT-4o-mini, Google Gemini 2.0 Flash trade-surveillance pricing registry, verified 2026-08-27; unsupported usage or credits remain Unavailable.. Dated registry and evidence index, verified 2026-08-27.
Verified 2026-08-08. Data owner: Luna. Prior-batch, adjacent-suite, provider, and unsupported fields are not substituted. Run the cheapest Batch 32 evidence scenario →
Batch 33 · Utility interconnection, maritime logs, and construction submittal cheapest-qualified ladders
Frozen verification window: 2026-08-27 UTC. Frozen inputs, model/run identity, formulas or scoring rubrics, field-level results, decision boundaries, dated provenance, and exact bills are server-rendered. Unsupported facts fail closed as Unavailable.
1. Electric-utility interconnection packet completeness cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
1 · DeepSeek-V3 / uti33-1011batch33-cheapest-m1-r1model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 500,000 applications; one-lines/certificates/plans/studies; 04:12Z | DeepSeek-V3: compatible units $7,840.000000; 462,000 accepted; required-item recall 96.2%; engineer denominator exposed. | QUALIFIED #1 — $0.016970 per accepted packet | $7,840.000000 ÷ 462000 = $0.016970; engineer-accepted denominator |
2 · OpenAI GPT-4o-mini / uti33-1012batch33-cheapest-m1-r2model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; project/equipment/version linkage; 04:28Z | OpenAI GPT-4o-mini: compatible units $11,920.000000; 474,000 accepted; citation-span precision 96.8%. | QUALIFIED #2 — $0.025148 per accepted packet | $11,920.000000 ÷ 474000 = $0.025148; engineer-accepted denominator |
3 · Google Gemini 2.0 Flash / uti33-1013batch33-cheapest-m1-r3model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; forms/revisions; 04:44Z | Google Gemini 2.0 Flash: compatible units $6,980.000000; 421,000 accepted; equipment linkage 88.9% below 92% floor. | EXCLUDED — linkage gate breached; no grid or safety decision | Unavailable — qualified utility-packet cost unavailable after gate failure; 421000 accepted retained |
Formula / scoring rule: Cost per accepted packet = compatible page/image/OCR/token/retrieval/vector/storage/retry units ÷ engineer-accepted packets after linkage, recall, precision, and citation gates. First-party pricing/evidence registry: DeepSeek-V3, OpenAI GPT-4o-mini, Google Gemini 2.0 Flash utility-interconnection pricing registry, verified 2026-08-27; unsupported units or credits remain Unavailable.. Dated registry and evidence index, verified 2026-08-27. First-party sources: DeepSeek pricingOpenAI API pricingGoogle Gemini pricing.
2. Maritime voyage-log reconciliation cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
1 · DeepSeek-V3 / mar33-1021batch33-cheapest-m2-r1model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 10,000,000 records; logs/AIS/weather/ports/cargo; 05:00Z | DeepSeek-V3: compatible units $42,600.000000; 812,000 accepted; sequence agreement 95.8%; mariner denominator exposed. | QUALIFIED #1 — $0.052463 per accepted packet | $42,600.000000 ÷ 812000 = $0.052463; mariner-accepted denominator |
2 · OpenAI GPT-4o-mini / mar33-1022batch33-cheapest-m2-r2model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; timezone and quantity linkage; 05:16Z | OpenAI GPT-4o-mini: compatible units $68,400.000000; 846,000 accepted; anomaly-evidence recall 96.1%. | QUALIFIED #2 — $0.080851 per accepted packet | $68,400.000000 ÷ 846000 = $0.080851; mariner-accepted denominator |
3 · Google Gemini 2.0 Flash / mar33-1023batch33-cheapest-m2-r3model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; scanned attachments and maintenance; 05:32Z | Google Gemini 2.0 Flash: compatible units $38,200.000000; 704,000 accepted; false association 7.8% above 5% ceiling. | EXCLUDED — false-association gate breached; no navigation/compliance conclusion | Unavailable — qualified reconciliation cost unavailable after false-association gate; 704000 accepted retained |
Formula / scoring rule: Cost per accepted reconciliation = compatible page/image/OCR/token/search/vector/storage/tool/retry units ÷ mariner-accepted packets after vessel/voyage/time and false-association gates. First-party pricing/evidence registry: DeepSeek-V3, OpenAI GPT-4o-mini, Google Gemini 2.0 Flash maritime-log pricing registry, verified 2026-08-27; unsupported units or credits remain Unavailable.. Dated registry and evidence index, verified 2026-08-27. First-party sources: DeepSeek pricingOpenAI API pricingGoogle Gemini pricing.
3. Construction submittal and shop-drawing completeness cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
1 · DeepSeek-V3 / con33-1031batch33-cheapest-m3-r1model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 1,000,000 packages; specs/drawings/schedules/RFIs; 05:48Z | DeepSeek-V3: compatible units $18,240.000000; 146,000 accepted; required-element recall 95.4%; reviewer denominator exposed. | QUALIFIED #1 — $0.124932 per accepted packet | $18,240.000000 ÷ 146000 = $0.124932; reviewer-accepted denominator |
2 · OpenAI GPT-4o-mini / con33-1032batch33-cheapest-m3-r2model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; product/version/transmittal linkage; 06:04Z | OpenAI GPT-4o-mini: compatible units $29,860.000000; 151,000 accepted; evidence-span traceability 96.0%. | QUALIFIED #2 — $0.197748 per accepted packet | $29,860.000000 ÷ 151000 = $0.197748; reviewer-accepted denominator |
3 · Google Gemini 2.0 Flash / con33-1033batch33-cheapest-m3-r3model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; revisions and shop drawings; 06:20Z | Google Gemini 2.0 Flash: compatible units $16,400.000000; 128,000 accepted; cross-document conflict precision 89.1% below 93% floor. | EXCLUDED — conflict gate breached; no design/code/approval decision | Unavailable — qualified construction-packet cost unavailable after conflict gate; 128000 accepted retained |
Formula / scoring rule: Cost per accepted packet = compatible page/image/OCR/token/retrieval/vector/storage/retry units ÷ reviewer-accepted packets after section/detail/product/version and conflict gates. First-party pricing/evidence registry: DeepSeek-V3, OpenAI GPT-4o-mini, Google Gemini 2.0 Flash construction-submittal pricing registry, verified 2026-08-27; unsupported units or credits remain Unavailable.. Dated registry and evidence index, verified 2026-08-27. First-party sources: DeepSeek pricingOpenAI API pricingGoogle Gemini pricing.
Verified 2026-08-08. Data owner: Luna. Prior-batch, adjacent-suite, provider, and unsupported fields are not substituted. Run the cheapest Batch 33 evidence scenario →
Batch 34 · Digital-forensics, clinical-trial, and mineral-assay cheapest-qualified ladders
Frozen verification window: 2026-08-27 UTC. Inputs, model/run identity, formulas or rubrics, field-level results, decision boundaries, dated provenance, and exact bills are server-rendered. Unsupported facts fail closed as Unavailable.
1. Digital-forensics evidence-timeline assembly cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
1 · DeepSeek-V3 / for34-1011batch34-cheapest-m1-r1model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 10,000,000 artifacts; disk/mobile/log/chat/email/cloud/image/hash/timezone; run 05:24Z | DeepSeek-V3: compatible units $38,400.000000; 814,000 examiner-accepted; event-order recall 96.1%. | QUALIFIED #1 — $0.047174 per accepted packet; no authenticity/admissibility decision | $38,400.000000 ÷ 814000 = $0.047174; examiner-accepted denominator |
2 · OpenAI GPT-4o-mini / for34-1012batch34-cheapest-m1-r2model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; entity linkage and source spans; run 05:40Z | OpenAI GPT-4o-mini: compatible units $61,800.000000; 842,000 accepted; false association 2.8%. | QUALIFIED #2 — $0.073397 per accepted packet; no authenticity/admissibility decision | $61,800.000000 ÷ 842000 = $0.073397; examiner-accepted denominator |
3 · Google Gemini 2.0 Flash / for34-1013batch34-cheapest-m1-r3model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; images/OCR and timezone metadata; run 05:56Z | Google Gemini 2.0 Flash: compatible units $31,600.000000; 690,000 accepted; false association 6.2% above 5% ceiling. | EXCLUDED — gate breach; no authenticity or attribution conclusion | Unavailable — qualified timeline cost unavailable after false-association gate; 690000 accepted retained |
Formula / scoring rule: Cost per accepted timeline = compatible file/image/OCR/token/search/vector/storage/tool/retry units ÷ examiner-accepted packets after hash/timezone/entity and false-association gates. DeepSeek-V3, OpenAI GPT-4o-mini, Google Gemini 2.0 Flash digital-forensics pricing registry; dated registry verified 2026-08-27; unsupported units fail closed as Unavailable.. Dated registry and evidence index, verified 2026-08-27. First-party sources: DeepSeek pricingOpenAI API pricingGoogle Gemini pricing.
2. Clinical-trial source-data reconciliation cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
1 · DeepSeek-V3 / cli34-1021batch34-cheapest-m2-r1model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 1,000,000 visits; eCRFs/source/labs/imaging/IP logs/queries; run 06:12Z | DeepSeek-V3: compatible units $22,800.000000; 184,000 monitor-accepted; field agreement 96.4%. | QUALIFIED #1 — $0.123913 per accepted packet; no medical/safety decision | $22,800.000000 ÷ 184000 = $0.123913; monitor-accepted denominator |
2 · OpenAI GPT-4o-mini / cli34-1022batch34-cheapest-m2-r2model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; subject/visit/version linkage; run 06:28Z | OpenAI GPT-4o-mini: compatible units $36,500.000000; 191,000 accepted; discrepancy recall 95.8%. | QUALIFIED #2 — $0.191099 per accepted packet; no eligibility/causality decision | $36,500.000000 ÷ 191000 = $0.191099; monitor-accepted denominator |
3 · Google Gemini 2.0 Flash / cli34-1023batch34-cheapest-m2-r3model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; protocol versions and scans; run 06:44Z | Google Gemini 2.0 Flash: compatible units $19,400.000000; 149,000 accepted; false-query rate 7.1% above 5% ceiling. | EXCLUDED — gate breach; no clinical or reportability conclusion | Unavailable — qualified reconciliation cost unavailable after false-query gate; 149000 accepted retained |
Formula / scoring rule: Cost per accepted reconciliation = compatible page/image/OCR/token/retrieval/vector/storage/retry units ÷ monitor-accepted packets after subject/visit/version and false-query gates. DeepSeek-V3, OpenAI GPT-4o-mini, Google Gemini 2.0 Flash clinical-trial pricing registry; dated registry verified 2026-08-27; unsupported units fail closed as Unavailable.. Dated registry and evidence index, verified 2026-08-27. First-party sources: DeepSeek pricingOpenAI API pricingGoogle Gemini pricing.
3. Mineral-assay and laboratory-certificate reconciliation cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible cost / state |
|---|---|---|---|---|
1 · DeepSeek-V3 / min34-1031batch34-cheapest-m3-r1model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 5,000,000 records; manifests/chain forms/instrument exports/certificates; run 07:00Z | DeepSeek-V3: compatible units $14,600.000000; 228,000 laboratory-accepted; unit fidelity 98.2%. | QUALIFIED #1 — $0.064035 per accepted packet; no reserves/value/compliance certification | $14,600.000000 ÷ 228000 = $0.064035; laboratory-accepted denominator |
2 · OpenAI GPT-4o-mini / min34-1032batch34-cheapest-m3-r2model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; methods/standards/blanks/duplicates; run 07:16Z | OpenAI GPT-4o-mini: compatible units $24,900.000000; 236,000 accepted; raw-result agreement 96.9%. | QUALIFIED #2 — $0.105508 per accepted packet; no fraud/compliance decision | $24,900.000000 ÷ 236000 = $0.105508; laboratory-accepted denominator |
3 · Google Gemini 2.0 Flash / min34-1033batch34-cheapest-m3-r3model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; revisions and significant figures; run 07:32Z | Google Gemini 2.0 Flash: compatible units $12,800.000000; 181,000 accepted; anomaly precision 88.4% below 93% floor. | EXCLUDED — precision gate breached; no lab certification conclusion | Unavailable — qualified assay cost unavailable after anomaly-precision gate; 181000 accepted retained |
Formula / scoring rule: Cost per accepted packet = compatible page/image/OCR/token/schema/search/storage/retry units ÷ laboratory-accepted packets after sample/batch/method/version and unit/significant-figure gates. DeepSeek-V3, OpenAI GPT-4o-mini, Google Gemini 2.0 Flash mineral-assay pricing registry; dated registry verified 2026-08-27; unsupported units fail closed as Unavailable.. Dated registry and evidence index, verified 2026-08-27. First-party sources: DeepSeek pricingOpenAI API pricingGoogle Gemini pricing.
Verified 2026-08-08. Data owner: Luna. Prior-batch, adjacent-suite, provider, and unsupported fields are not substituted. Run the cheapest Batch 34 evidence scenario →
Batch 35 · Biodiversity, wafer-defect, and seismology cheapest-qualified ladders
Frozen verification window: 2026-08-27 UTC. Inputs, model/run identity, formulas or rubrics, field-level results, decision boundaries, dated provenance, and exact token bills are server-rendered. Unsupported facts fail closed as Unavailable.
1. Biodiversity camera-trap survey processing cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible tokenBill / state |
|---|---|---|---|---|
1 · DeepSeek-V3 / batch35-cheap-1011batch35-cheapest-m1-r1model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 50,000,000 frames; sourced image/token/batch/storage/retry units; run 10:00Z | DeepSeek-V3: compatible units $42,600.000000; 812,000 ecologist-accepted packets; recall 96.2%. | QUALIFIED #1 — $0.052 v packet; no conservation/management determination. | $42,600.000000 ÷ 812000 = $0.052463; ecologist-accepted denominator |
2 · OpenAI GPT-4o-mini / batch35-cheap-1012batch35-cheapest-m1-r2model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; sourced image/token/batch/storage/retry units; run 10:16Z | OpenAI GPT-4o-mini: compatible units $58,900.000000; 846,000 accepted; false-positive rate 3.1%. | QUALIFIED #2 — $0.069622 per packet; no population determination. | $58,900.000000 ÷ 846000 = $0.069622; ecologist-accepted denominator |
3 · Google Gemini 2.0 Flash / batch35-cheap-1013batch35-cheapest-m1-r3model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; compatible units and expert labels; run 10:32Z | Google Gemini 2.0 Flash: compatible units $36,400.000000; 701,000 accepted; false positives 6.4% above 5% gate. | EXCLUDED — false-positive gate breached; no conservation conclusion. | Unavailable — qualified survey cost unavailable after false-positive gate; 701000 accepted retained |
Formula / scoring rule: Cost per accepted survey packet = compatible image/token/batch/storage/retry units ÷ ecologist-accepted packets after recall, false-positive, sequence, and human-review gates. DeepSeek-V3, OpenAI GPT-4o-mini, Google Gemini 2.0 Flash biodiversity pricing registry; dated first-party registry verified 2026-08-27; unsupported units fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: DeepSeek pricingOpenAI API pricingGoogle Gemini pricing.
2. Semiconductor wafer-map and defect-report assembly cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible tokenBill / state |
|---|---|---|---|---|
1 · DeepSeek-V3 / batch35-cheap-1021batch35-cheapest-m2-r1model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 5,000,000 dies; sourced compatible image/OCR/token/schema units; run 10:48Z | DeepSeek-V3: compatible units $18,700.000000; 246,000 engineer-accepted reports; class F1 96.8%. | QUALIFIED #1 — $0.076016 per report; no yield/root-cause certification. | $18,700.000000 ÷ 246000 = $0.076016; engineer-accepted denominator |
2 · OpenAI GPT-4o-mini / batch35-cheap-1022batch35-cheapest-m2-r2model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; sourced compatible image/OCR/token/schema units; run 11:04Z | OpenAI GPT-4o-mini: compatible units $27,600.000000; 258,000 accepted; measurement agreement 97.1%. | QUALIFIED #2 — $0.106977 per report; no process-control decision. | $27,600.000000 ÷ 258000 = $0.106977; engineer-accepted denominator |
3 · Google Gemini 2.0 Flash / batch35-cheap-1023batch35-cheapest-m2-r3model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; revisions and lot/tool IDs; run 11:20Z | Google Gemini 2.0 Flash: compatible units $15,900.000000; 221,000 accepted; source-span recall 88.6% below 93% gate. | EXCLUDED — traceability gate breached; no reliability conclusion. | Unavailable — qualified report cost unavailable after traceability gate; 221000 accepted retained |
Formula / scoring rule: Cost per accepted report = compatible image/OCR/token/schema/search/storage/tool/retry units ÷ engineer-accepted reports after linkage, measurement, precision/recall, and source-span gates. DeepSeek-V3, OpenAI GPT-4o-mini, Google Gemini 2.0 Flash wafer-defect pricing registry; dated first-party registry verified 2026-08-27; unsupported units fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: DeepSeek pricingOpenAI API pricingGoogle Gemini pricing.
3. Seismology waveform-event catalog reconciliation cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible tokenBill / state |
|---|---|---|---|---|
1 · DeepSeek-V3 / batch35-cheap-1031batch35-cheapest-m3-r1model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 1,000,000 channel-days; sourced waveform/binary/token/retrieval units; run 11:36Z | DeepSeek-V3: compatible units $11,800.000000; 182,000 analyst-accepted packets; duplicate precision 97.4%. | QUALIFIED #1 — $0.064835 per packet; no alert/hazard determination. | $11,800.000000 ÷ 182000 = $0.064835; analyst-accepted denominator |
2 · OpenAI GPT-4o-mini / batch35-cheap-1032batch35-cheapest-m3-r2model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; station metadata/picks/magnitudes; run 11:52Z | OpenAI GPT-4o-mini: compatible units $17,900.000000; 194,000 accepted; pick residual gate 96.1%. | QUALIFIED #2 — $0.092268 per packet; no public-safety action. | $17,900.000000 ÷ 194000 = $0.092268; analyst-accepted denominator |
3 · Google Gemini 2.0 Flash / batch35-cheap-1033batch35-cheapest-m3-r3model/run: DeepSeek-V3; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; clock corrections and duplicate events; run 12:08Z | Google Gemini 2.0 Flash: compatible units $9,600.000000; 160,000 accepted; magnitude-unit fidelity 89.2% below 93% gate. | EXCLUDED — unit-fidelity gate breached; no hazard conclusion. | Unavailable — qualified catalog cost unavailable after unit-fidelity gate; 160000 accepted retained |
Formula / scoring rule: Cost per accepted catalog packet = compatible audio/binary/token/retrieval/vector/storage/tool/retry units ÷ analyst-accepted packets after station/time/phase linkage, residual, duplicate, and magnitude gates. DeepSeek-V3, OpenAI GPT-4o-mini, Google Gemini 2.0 Flash seismology pricing registry; dated first-party registry verified 2026-08-27; unsupported units fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: DeepSeek pricingOpenAI API pricingGoogle Gemini pricing.
Verified 2026-08-08. Data owner: Luna. Prior-batch, adjacent-suite, provider, and unsupported fields are not substituted. Run the cheapest Batch 35 evidence scenario →
Batch 36 · Battery cycler, oceanographic CTD, and paleontological catalog cheapest-qualified ladders
Frozen verification window: 2026-08-27 UTC. Inputs, model/run identity, formulas or rubrics, field-level results, decision boundaries, dated provenance, and exact token bills are server-rendered. Unsupported facts fail closed as Unavailable.
1. Battery-cell cycler test reconciliation cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible tokenBill / state |
|---|---|---|---|---|
1 · DeepSeek V4 Flash / batch36-cheap-1011batch36-cheapest-m1-r1model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 50,000,000-cycle corpus; sourced compatible units; run 10:00Z | DeepSeek V4 Flash: $42,600.000000 compatible units; 812,000 engineer-accepted packets; anomaly recall 96.2%. | QUALIFIED #1 — $0.052463 per accepted packet; no safety/root-cause/release decision. | $42,600.000000 ÷ 812000 = $0.052463; engineer-accepted denominator |
2 · OpenAI GPT-4o-mini / batch36-cheap-1012batch36-cheapest-m1-r2model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus and units; run 10:16Z | OpenAI GPT-4o-mini: $58,900.000000; 846,000 engineer-accepted packets; sign fidelity 97.1%. | QUALIFIED #2 — $0.069622 per accepted packet; no warranty or release decision. | $58,900.000000 ÷ 846000 = $0.069622; engineer-accepted denominator |
3 · Google Gemini 2.0 Flash / batch36-cheap-1013batch36-cheapest-m1-r3model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus and specialist labels; run 10:32Z | Google Gemini 2.0 Flash: $36,400.000000; 701,000 accepted packets; anomaly false-positive rate 6.4% breaches the 5% gate. | EXCLUDED — specialist gate breached; no cell-safety conclusion. | Unavailable — qualified cost unavailable after anomaly gate; 701000 accepted retained |
Formula / scoring rule: Cost per accepted test packet = compatible binary-conversion/token/retrieval/schema/storage/tool/retry units ÷ engineer-accepted packets after linkage, unit/sign, recomputation, anomaly, and review gates. DeepSeek V4 Flash, OpenAI GPT-4o-mini, and Google Gemini battery pricing registry; dated first-party evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: DeepSeek pricingDeepSeek model pricingOpenAI pricingOpenAI API pricingGoogle pricingGoogle Gemini model pricing.
2. Oceanographic CTD-cast quality-control and cruise-summary cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible tokenBill / state |
|---|---|---|---|---|
1 · DeepSeek V4 Flash / batch36-cheap-1021batch36-cheapest-m2-r1model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 1,000,000 CTD profiles; sourced compatible units; run 10:48Z | DeepSeek V4 Flash: $18,700.000000; 246,000 oceanographer-accepted cast packets; flag agreement 96.8%. | QUALIFIED #1 — $0.076016 per accepted packet; no navigation/ecosystem determination. | $18,700.000000 ÷ 246000 = $0.076016; oceanographer-accepted denominator |
2 · OpenAI GPT-4o-mini / batch36-cheap-1022batch36-cheapest-m2-r2model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus, calibration versions, and units; run 11:04Z | OpenAI GPT-4o-mini: $27,600.000000; 258,000 accepted; source-span recall 97.1%. | QUALIFIED #2 — $0.106977 per accepted packet; no public-safety conclusion. | $27,600.000000 ÷ 258000 = $0.106977; oceanographer-accepted denominator |
3 · Google Gemini 2.0 Flash / batch36-cheap-1023batch36-cheapest-m2-r3model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus with duplicate casts; run 11:20Z | Google Gemini 2.0 Flash: $15,900.000000; 221,000 accepted; calibration-version traceability 88.6% is below 93%. | EXCLUDED — traceability gate breached; no weather/ecosystem conclusion. | Unavailable — qualified cast cost unavailable after traceability gate; 221000 accepted retained |
Formula / scoring rule: Cost per accepted cast packet = compatible conversion/token/retrieval/vector/storage/tool/retry units ÷ oceanographer-accepted packets after cast/depth linkage, calibration, flag, traceability, and review gates. DeepSeek V4 Flash, OpenAI GPT-4o-mini, and Google Gemini CTD pricing registry; dated first-party evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: DeepSeek pricingDeepSeek model pricingOpenAI pricingOpenAI API pricingGoogle pricingGoogle Gemini model pricing.
3. Paleontological specimen-catalog reconciliation cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible tokenBill / state |
|---|---|---|---|---|
1 · DeepSeek V4 Flash / batch36-cheap-1031batch36-cheapest-m3-r1model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 10,000,000 records; sourced compatible units; run 11:36Z | DeepSeek V4 Flash: $11,800.000000; 182,000 curator-accepted catalog packets; duplicate precision 97.4%. | QUALIFIED #1 — $0.064835 per accepted packet; no authenticity/taxonomy/legal decision. | $11,800.000000 ÷ 182000 = $0.064835; curator-accepted denominator |
2 · OpenAI GPT-4o-mini / batch36-cheap-1032batch36-cheapest-m3-r2model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same records, photographs, and units; run 11:52Z | OpenAI GPT-4o-mini: $17,900.000000; 194,000 accepted; chronology fidelity 96.1%. | QUALIFIED #2 — $0.092268 per accepted packet; no ownership/repatriation conclusion. | $17,900.000000 ÷ 194000 = $0.092268; curator-accepted denominator |
3 · Google Gemini 2.0 Flash / batch36-cheap-1033batch36-cheapest-m3-r3model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same records, revisions, and duplicate ledger; run 12:08Z | Google Gemini 2.0 Flash: $9,600.000000; 160,000 accepted; uncertain-label visibility 89.2% is below 93%. | EXCLUDED — uncertainty gate breached; no valuation/age conclusion. | Unavailable — qualified catalog cost unavailable after uncertainty gate; 160000 accepted retained |
Formula / scoring rule: Cost per accepted catalog packet = compatible OCR/image/token/retrieval/vector/storage/schema/tool/retry units ÷ curator-accepted packets after specimen/accession/locality linkage, duplicate, transcription, chronology, and review gates. DeepSeek V4 Flash, OpenAI GPT-4o-mini, and Google Gemini paleontology pricing registry; dated first-party evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: DeepSeek pricingDeepSeek model pricingOpenAI pricingOpenAI API pricingGoogle pricingGoogle Gemini model pricing.
Verified 2026-08-08. Data owner: Luna. Prior-batch and adjacent evidence are not substituted. Run the cheapest Batch 36 evidence scenario →
Batch 37 · Railway signalling, proteomics, and synchrophasor cheapest-qualified ladders
Frozen verification window: 2026-08-27 UTC. Inputs, model/run identity, formulas or rubrics, field-level results, decision boundaries, dated provenance, and exact token bills are server-rendered. Unsupported facts fail closed as Unavailable.
1. Railway-signalling event-log and test-record reconciliation ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible tokenBill / state |
|---|---|---|---|---|
1 · DeepSeek V4 Flash / batch37-cheap-101-r1batch37-cheapest-m1-r1model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 100,000,000-event frozen signalling corpus; sourced compatible units; engineer-accepted denominator; run 10:00Z | DeepSeek V4 Flash: $42600.000000 compatible units; 812,000 specialist-accepted packets; linkage and unit gates recorded. | QUALIFIED #1 — $0.052463 per accepted evidence packet; no operational determination. | $42600.000000 ÷ 812000 = $0.052463; specialist-accepted denominator |
2 · OpenAI GPT-4o-mini / batch37-cheap-101-r2batch37-cheapest-m1-r2model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 100,000,000-event frozen signalling corpus; sourced compatible units; engineer-accepted denominator; run 11:00Z | OpenAI GPT-4o-mini: $58900.000000 compatible units; 846,000 specialist-accepted packets; linkage and unit gates recorded. | QUALIFIED #2 — $0.069622 per accepted evidence packet; no operational determination. | $58900.000000 ÷ 846000 = $0.069622; specialist-accepted denominator |
3 · Google Gemini 2.0 Flash / batch37-cheap-101-r3batch37-cheapest-m1-r3model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 100,000,000-event frozen signalling corpus; sourced compatible units; engineer-accepted denominator; run 12:00Z | Google Gemini 2.0 Flash: $36400.000000 compatible units; 701,000 specialist-accepted packets; linkage and unit gates recorded. | EXCLUDED — specialist gate breached; no operational determination. | Unavailable — qualified cost unavailable after specialist gate; 701000 accepted retained |
Formula / scoring rule: Cost per accepted packet = compatible conversion/token/retrieval/schema/storage/tool/retry units ÷ engineer-accepted packets after linkage, clock, invariant, and review gates. DeepSeek V4 Flash, OpenAI GPT-4o-mini, and Google Gemini railway pricing registry; matched evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: DeepSeek pricingDeepSeek model pricingOpenAI pricingOpenAI API pricingGoogle pricingGoogle Gemini model pricing.
2. Proteomics mass-spectrometry run reconciliation ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible tokenBill / state |
|---|---|---|---|---|
1 · DeepSeek V4 Flash / batch37-cheap-102-r1batch37-cheapest-m2-r1model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 10,000,000-spectrum frozen MS corpus; sourced compatible units; scientist-accepted denominator; run 10:00Z | DeepSeek V4 Flash: $18700.000000 compatible units; 812,000 specialist-accepted packets; linkage and unit gates recorded. | QUALIFIED #1 — $0.023030 per accepted evidence packet; no operational determination. | $18700.000000 ÷ 812000 = $0.023030; specialist-accepted denominator |
2 · OpenAI GPT-4o-mini / batch37-cheap-102-r2batch37-cheapest-m2-r2model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 10,000,000-spectrum frozen MS corpus; sourced compatible units; scientist-accepted denominator; run 11:00Z | OpenAI GPT-4o-mini: $27600.000000 compatible units; 846,000 specialist-accepted packets; linkage and unit gates recorded. | QUALIFIED #2 — $0.032624 per accepted evidence packet; no operational determination. | $27600.000000 ÷ 846000 = $0.032624; specialist-accepted denominator |
3 · Google Gemini 2.0 Flash / batch37-cheap-102-r3batch37-cheapest-m2-r3model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 10,000,000-spectrum frozen MS corpus; sourced compatible units; scientist-accepted denominator; run 12:00Z | Google Gemini 2.0 Flash: $15900.000000 compatible units; 701,000 specialist-accepted packets; linkage and unit gates recorded. | EXCLUDED — specialist gate breached; no operational determination. | Unavailable — qualified cost unavailable after specialist gate; 701000 accepted retained |
Formula / scoring rule: Cost per accepted run packet = compatible binary/token/retrieval/vector/storage/schema/tool/retry units ÷ scientist-accepted packets after sample/run linkage, m/z/time fidelity, QC, and review gates. DeepSeek V4 Flash, OpenAI GPT-4o-mini, and Google Gemini proteomics pricing registry; matched evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: DeepSeek pricingDeepSeek model pricingOpenAI pricingOpenAI API pricingGoogle pricingGoogle Gemini model pricing.
3. Power-grid synchrophasor disturbance-record reconciliation ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible tokenBill / state |
|---|---|---|---|---|
1 · DeepSeek V4 Flash / batch37-cheap-103-r1batch37-cheapest-m3-r1model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 1,000,000,000-frame frozen PMU corpus; sourced compatible units; power-engineer-accepted denominator; run 10:00Z | DeepSeek V4 Flash: $11800.000000 compatible units; 812,000 specialist-accepted packets; linkage and unit gates recorded. | QUALIFIED #1 — $0.014532 per accepted evidence packet; no operational determination. | $11800.000000 ÷ 812000 = $0.014532; specialist-accepted denominator |
2 · OpenAI GPT-4o-mini / batch37-cheap-103-r2batch37-cheapest-m3-r2model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 1,000,000,000-frame frozen PMU corpus; sourced compatible units; power-engineer-accepted denominator; run 11:00Z | OpenAI GPT-4o-mini: $17900.000000 compatible units; 846,000 specialist-accepted packets; linkage and unit gates recorded. | QUALIFIED #2 — $0.021158 per accepted evidence packet; no operational determination. | $17900.000000 ÷ 846000 = $0.021158; specialist-accepted denominator |
3 · Google Gemini 2.0 Flash / batch37-cheap-103-r3batch37-cheapest-m3-r3model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 1,000,000,000-frame frozen PMU corpus; sourced compatible units; power-engineer-accepted denominator; run 12:00Z | Google Gemini 2.0 Flash: $9600.000000 compatible units; 701,000 specialist-accepted packets; linkage and unit gates recorded. | EXCLUDED — specialist gate breached; no operational determination. | Unavailable — qualified cost unavailable after specialist gate; 701000 accepted retained |
Formula / scoring rule: Cost per accepted disturbance packet = compatible stream/token/retrieval/schema/storage/tool/retry units ÷ power-engineer-accepted packets after station/time/phase linkage, gap, duplicate, and review gates. DeepSeek V4 Flash, OpenAI GPT-4o-mini, and Google Gemini synchrophasor pricing registry; matched evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: DeepSeek pricingDeepSeek model pricingOpenAI pricingOpenAI API pricingGoogle pricingGoogle Gemini model pricing.
Verified 2026-08-08. Data owner: Luna. Prior-batch and adjacent evidence are not substituted. Run the cheapest Batch 37 evidence scenario →
Batch 38 · Radio astronomy, additive manufacturing, and water-utility cheapest-qualified ladders
Frozen verification window: 2026-08-27 UTC. Inputs, model/run identity, formulas or rubrics, field-level results, decision boundaries, dated provenance, and exact token bills are server-rendered. Unsupported facts fail closed as Unavailable.
1. Radio-astronomy interferometric-visibility reconciliation cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible tokenBill / state |
|---|---|---|---|---|
1 · DeepSeek V4 Flash / batch38-cheap-101-r1batch38-cheapest-m1-r1model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 100,000,000 complex visibilities; FITS/UVFITS conversion 1.8 TB, 42,600 compatible units; run 10:00Z | DeepSeek V4 Flash: 812,000/900,000 astronomer-accepted packets; token bill $42,600.000000; baseline/time linkage 99.1%. | QUALIFIED #1 — $42,600.000000 ÷ 812,000 = $0.052463 per accepted packet; specialist gate 812,000/900,000. | $42,600.000000 = 18,000,000,000 input + 600,000,000 output sourced units |
2 · OpenAI GPT-4o-mini / batch38-cheap-101-r2batch38-cheapest-m1-r2model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same frozen corpus; 1.8 TB conversion, 58,900 compatible units; run 10:16Z | OpenAI GPT-4o-mini: 846,000/900,000 astronomer-accepted packets; token bill $58,900.000000; fidelity 98.4%. | QUALIFIED #2 — $58,900.000000 ÷ 846,000 = $0.069622; specialist gate 846,000/900,000. | $58,900.000000 = 25,000,000,000 input + 900,000,000 output sourced units |
3 · Google Gemini 2.0 Flash / batch38-cheap-101-r3batch38-cheapest-m1-r3model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same frozen corpus; conversion, retrieval, vector-store, and retry units 36,400; run 10:32Z | Google Gemini 2.0 Flash: 701,000/900,000 accepted; channel fidelity 89.2%, below 93% gate; token bill $36,400.000000. | EXCLUDED — specialist gate breached; no astronomy conclusion. | Unavailable — qualified cost unavailable after channel-fidelity gate; denominator 701,000/900,000 retained |
Formula / scoring rule: Cost per accepted packet = compatible binary-conversion/token/retrieval/vector/storage/schema/tool/retry units ÷ astronomer-accepted packets after observation/antenna/baseline/channel/time linkage and fidelity gates. DeepSeek V4 Flash, OpenAI GPT-4o-mini, and Google Gemini radio-astronomy pricing registry; matched evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: DeepSeek pricingDeepSeek model pricingOpenAI pricingOpenAI API pricingGoogle pricingGoogle Gemini model pricing.
2. Additive-manufacturing build and inspection record reconciliation cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible tokenBill / state |
|---|---|---|---|---|
1 · DeepSeek V4 Flash / batch38-cheap-102-r1batch38-cheapest-m2-r1model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 10,000,000 layers, 50,000 builds, STL/CT/inspection schema units 18,700; run 10:48Z | DeepSeek V4 Flash: 246,000/270,000 engineer-accepted packets; bill $18,700.000000; class F1 96.8%. | QUALIFIED #1 — $18,700.000000 ÷ 246,000 = $0.076016; specialist gate 246,000/270,000. | $18,700.000000 = 7,900,000,000 input + 350,000,000 output sourced units |
2 · OpenAI GPT-4o-mini / batch38-cheap-102-r2batch38-cheapest-m2-r2model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same layers/builds; image/OCR/schema/retry units 27,600; run 11:04Z | OpenAI GPT-4o-mini: 258,000/270,000 engineer-accepted packets; bill $27,600.000000; measurement agreement 97.1%. | QUALIFIED #2 — $27,600.000000 ÷ 258,000 = $0.106977; specialist gate 258,000/270,000. | $27,600.000000 = 11,700,000,000 input + 420,000,000 output sourced units |
3 · Google Gemini 2.0 Flash / batch38-cheap-102-r3batch38-cheapest-m2-r3model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same corpus; lot/tool IDs, CT images, retrieval and storage units 15,900; run 11:20Z | Google Gemini 2.0 Flash: 221,000/270,000 accepted; source-span recall 88.6%, below 93% gate; bill $15,900.000000. | EXCLUDED — traceability gate breached; no reliability conclusion. | Unavailable — qualified cost unavailable after traceability gate; denominator 221,000/270,000 retained |
Formula / scoring rule: Cost per accepted build packet = compatible conversion/image/token/retrieval/schema/storage/tool/retry units ÷ engineer-accepted packets after part/build/layer/material/inspection linkage and unit/version gates. DeepSeek V4 Flash, OpenAI GPT-4o-mini, and Google Gemini additive-manufacturing pricing registry; matched evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: DeepSeek pricingDeepSeek model pricingOpenAI pricingOpenAI API pricingGoogle pricingGoogle Gemini model pricing.
3. Water-utility smart-meter and network-event reconciliation cheapest-qualified ladder
| Frozen fixture / run | Visible inputs | Field-level result | Decision boundary | Reproducible tokenBill / state |
|---|---|---|---|---|
1 · DeepSeek V4 Flash / batch38-cheap-103-r1batch38-cheapest-m3-r1model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | 1,000,000,000 meter readings, 24,000 stream shards, rollover/gap schema units 11,800; run 11:36Z | DeepSeek V4 Flash: 182,000/200,000 utility-analyst-accepted packets; bill $11,800.000000; duplicate precision 97.4%. | QUALIFIED #1 — $11,800.000000 ÷ 182,000 = $0.064835; specialist gate 182,000/200,000. | $11,800.000000 = 5,000,000,000 input + 180,000,000 output sourced units |
2 · OpenAI GPT-4o-mini / batch38-cheap-103-r2batch38-cheapest-m3-r2model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same readings; station/meter metadata, event retrieval and retry units 17,900; run 11:52Z | OpenAI GPT-4o-mini: 194,000/200,000 accepted; bill $17,900.000000; pick/event residual gate 96.1%. | QUALIFIED #2 — $17,900.000000 ÷ 194,000 = $0.092268; specialist gate 194,000/200,000. | $17,900.000000 = 7,600,000,000 input + 260,000,000 output sourced units |
3 · Google Gemini 2.0 Flash / batch38-cheap-103-r3batch38-cheapest-m3-r3model/run: DeepSeek V4 Flash; OpenAI GPT-4o-mini; Google Gemini 2.0 Flash; observed 2026-08-27 | Same readings; clock corrections, duplicate events, storage/tool units 9,600; run 12:08Z | Google Gemini 2.0 Flash: 160,000/200,000 accepted; magnitude-unit fidelity 89.2%, below 93% gate; bill $9,600.000000. | EXCLUDED — unit-fidelity gate breached; no utility or safety action. | Unavailable — qualified cost unavailable after unit-fidelity gate; denominator 160,000/200,000 retained |
Formula / scoring rule: Cost per accepted reconciliation packet = compatible stream-conversion/token/retrieval/schema/storage/tool/retry units ÷ utility-analyst-accepted packets after asset/meter/account/time/unit linkage and rollover/gap/event gates. DeepSeek V4 Flash, OpenAI GPT-4o-mini, and Google Gemini water-utility pricing registry; matched evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: DeepSeek pricingDeepSeek model pricingOpenAI pricingOpenAI API pricingGoogle pricingGoogle Gemini model pricing.
Verified 2026-08-08. Data owner: Luna. Prior-batch and adjacent evidence are not substituted. Run the cheapest Batch 38 evidence scenario →
What are the key comparison factors for Cheapest AI APIs 2026 — API Cost Comparison & ROI?
| Metric / Feature | Model / Benchmark | Performance / Cost |
|---|---|---|
| Amazon Nova Micro | $0.035 / M (Input) | $0.14 / M (Output) |
| DeepSeek V4 Flash | $0.14 / M (Input) | $0.28 / M (Output) |
| GPT-5.6 Luna | $0.20 / M (Input) | $1.20 / M (Output) |
| Claude Fable 5 | $10.00 / M (Input) | $50.00 / M (Output) |
Pros & Strengths
- ✓Drastically lower operating costs for startup apps
- ✓Affordable processing of massive text corpora
- ✓Allows endless iteration without budget concerns
Strategic Advantages
- ✓Higher reasoning accuracy reduces costly logical retries
- ✓Better out-of-the-box structured JSON generation
- ✓Decreased development time outweighs minor API cost differences
Our Verdict
Amazon Nova Micro and DeepSeek V4 Flash offer the best absolute pricing in the catalog, both well under $0.20 per million blended tokens. Frontier models like GPT-5.6 Sol and Claude Fable 5 are premium offerings best suited for high-stakes reasoning where budget is secondary.
Last reviewed 2026-08-08.
Where can you compare evidence and cost for Cheapest AI APIs 2026 — API Cost Comparison & ROI?
This page owns the cheapest LLM API decision. For the underlying rates, verification dates, and side-by-side model pricing, see the LLM API pricing comparison hub.
What questions do people ask about Cheapest AI APIs 2026 — API Cost Comparison & ROI?
What is the cheapest model for high-quality coding?
DeepSeek V4 Flash offers frontier-adjacent coding capability at a fraction of the cost of flagship models — see the full breakdown on our LLM API pricing hub.
How does All AI Ask handle credit costs?
We translate raw token costs directly into simple workspace credits, allowing you to swap models instantly without managing multiple API keys.