← All providers

Cerebras API Pricing, Models & Rate Limits (2026)

Cerebras hosts other labs' open-weight models — gpt-oss-120b and Z.ai's GLM 4.7 — on its wafer-scale inference hardware. Like Groq, the product is speed, not training; Cerebras' wafer-scale engine is the fastest hosting option on this site for the models it serves.

How much does the Cerebras API cost?

Cerebras API pricing is usage-based and varies by model and token volume. Use the current model table and provider documentation to estimate a workload before committing to production.

Verified — source

For cross-provider quota units and fixed-workload capacity, see the LLM API rate-limit comparison; this page remains the authoritative owner for Cerebras provider facts.

Cerebras product vs API

Cerebras consumer access and API billing are separate surfaces; check the provider documentation for current account terms.

Three decisions unique to Cerebras

Cerebras current-model price mechanics

Current modelInputCached inputOutputBatchVerified
GPT-OSS 120B (Cerebras)$0.350/MUnavailable — no model cache rate$0.750/MUnavailable2026-06-14
Qwen 3.8 27B (Cerebras)$0.990/MUnavailable — no model cache rate$1.490/MUnavailable2026-10-01
GLM 4.7 (Cerebras)$2.250/MUnavailable — no model cache rate$2.750/MUnavailable2026-06-14

Messages API vs OpenAI compatibility map

ChoiceDecision ruleEvidence
BillingChatGPT plan never includes API creditsSeparate metered API account
Input/cached/outputToken prices are model rowsUse calculator for workload totals
Limits/authPer-model requests/minute and tokens/minute caps by tier · Bearer API keyVerify before production

Adoption map: what is documented versus unavailable

DimensionRecorded valueDecision consequence
Try Cerebras side by side →

Verified 2026-08-14. dated provider pricing/source →

Current models
3
Legacy models
0
Price range /M
$0.45–$2.38
Max context
256K
Median tok/s
2215
Next retirement
—

Cerebras model pricing

Compare Cerebras models by input, output, and blended token cost below.

ModelInput /MOutput /MBlended /M
GPT-OSS 120B (Cerebras)$0.35$0.75$0.45
Qwen 3.8 27B (Cerebras)$0.99$1.49$1.11
GLM 4.7 (Cerebras)$2.25$2.75$2.38

Pricing values are registry-backed and were most recently verified on 2026-10-01. Sources: https://www.cerebras.ai/pricing. Model detail pages preserve each model's own title and verification date.

* Blended comparison assumes 3 input tokens for every output token; it is not the provider's billing unit.

Speed

Fastest measured Cerebras model is GPT-OSS 120B (Cerebras) at 2450 tokens/sec (90ms TTFT), median across measured Cerebras models is 2215 tokens/sec. See the full speed benchmark methodology.

Best for

Muse Spark 1.3 Contributor is our pick for CodingMuse Spark 1.3 Contributor is our pick for Structured Data ExtractionMuse Spark 1.3 Contributor is our pick for Writing & ContentGemini 3.1 Pro is our pick for Math & ReasoningMuse Spark 1.3 Contributor is our pick for Agents & Tool Use

Related Cerebras pages

Cerebras alternatives →All LLM API pricing →Cerebras speed benchmarks →

Build with Cerebras

Cerebras rate limits →

Cerebras implementation details

Verified 2026-08-14 against source.

Cerebras publishes its current authentication, limits, and data-handling details in the linked documentation.

OpenAI-compatibleYes
API base URLhttps://api.cerebras.ai/v1
Auth modelBearer API key
Prompt cachingNot documented
Batch discountNot documented
Free tierFree tier with a daily token cap
Free-tier limitsFree access is subject to a daily token cap and account quota.
Free-tier expiryNot published
Rate-limit modelPer-model requests/minute and tokens/minute caps by tier
Data residencyUS
Trains on API dataNot documented
SLA publishedNo
DocsOfficial pricingStatus pageFree-tier terms

Switching to and from Cerebras

The closest parity-aware alternative to GLM 4.7 (Cerebras) ($2.38/M) outside Cerebras is GLM-5.3 Flash ($0.24/M, -90%) — a config migration. Biggest gap: loses documented data-residency options.
The closest parity-aware alternative to Qwen 3.8 27B (Cerebras) ($1.11/M) outside Cerebras is Mistral Small 3.1 ($0.26/M, -76.5%) — a config migration.
Full Cerebras alternatives comparison →

Calling Cerebras through All AI Ask

curl https://allaiask.com/api/v1/chat \
  -H "Authorization: Bearer $ALLAIASK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "cerebras-gpt-oss-120b", "messages": [{"role": "user", "content": "Hello"}]}'

FAQ

Is Cerebras OpenAI-compatible?

Yes — Cerebras's API base (https://api.cerebras.ai/v1) accepts the OpenAI SDK request/response shape, so existing OpenAI client code works with a base-URL and key swap.

Does Cerebras support prompt caching?

Not documented as of 2026-08-14 — we did not find a published prompt-caching feature for Cerebras. If that changes, this page updates.

Does Cerebras have a free tier?

Yes — Free tier with a daily token cap. Free access is subject to a daily token cap and account quota.

How much does the Cerebras API cost?

Current Cerebras models range from $0.45 to $2.38 per million blended tokens (3:1 input:output). Full per-model pricing is in the table below.

Where is Cerebras API data hosted?

US

Cerebras hosted identity, quota envelope, and claim-to-observation ledger

Evidence review · verified .

Cerebras hosted-identity register

Frozen cerebras fixture — Catalog rows, revisions, alternate hosts, precision, tokenizer, and lifecycle; first-party evidence checked 2026-08-14; unresolved joins render Unavailable.

Method: every calculation uses only the joined fields shown in its row; missing joins fail closed. First-party source: Cerebras Inference documentation. Verified: 2026-08-14.

FixtureNamed inputsObservationFormula / boundaryDecision
current catalog rowowner=Meta; endpoint ID=llama-3.3-70b; revision=alias; artifact=Unavailable; precision=Unavailable; lifecycle=currentCerebras endpoint and owner join; precision and artifact visibility remain Unavailable.identity complete = owner ∧ endpoint ID ∧ revision ∧ artifact ∧ precision ∧ tokenizer/template
Boundary: Catalog identity does not license result transfer from another host.		UNAVAILABLE — precision/artifact missing.
frozen revision change + same-name alternate hostCerebras revision=2026-08; alternate host revision=2026-07; model name=same; tokenizer=Unavailable; dependent fixtures=6Same name masks a revision mismatch; dependent fixtures remain host-specific.transfer eligible = exact host ∧ revision ∧ tokenizer/template ∧ fixture join
Boundary: Cerebras evidence cannot stand in for Groq evidence.		FAIL — revision/host mismatch.
owner update + undocumented precisionowner=Qwen; endpoint=joined; owner evidence=2026-08-14; precision=Unavailable; lifecycle=active; dependent workloads=9Owner update is dated; undocumented precision blocks a hardware-to-artifact inference.result-transfer = 1 only when declared identity fields all join
Boundary: Hardware claim cannot fill a missing precision field.		UNAVAILABLE — precision remains unknown.

Quota-and-token-envelope admission board

Frozen cerebras fixture — Account/model tier, request/token/time buckets, reserves, and scoped admission; first-party evidence checked 2026-08-14; unresolved joins render Unavailable.

Method: every calculation uses only the joined fields shown in its row; missing joins fail closed. First-party source: Cerebras Inference documentation. Verified: 2026-08-14.

FixtureNamed inputsObservationFormula / boundaryDecision
100 short chatstier=free; model=joined; request cap=100; token cap=Unavailable; reserve=8K; admitted=100; deferred=0Request cap admits the declared 100; token cap remains unknown and does not become unlimited.admitted = min(requests, documented request cap, documented token envelope)
Boundary: Free-tier evidence is not provider-wide paid-tier evidence.		PASS WITH UNAVAILABLE TOKEN CAP — scoped.
eight 64K prompts + four long outputstier=paid; prompt reserve=512K; output reserve=32K; time bucket=Unavailable; admitted=Unavailable; retry=headerPrompt/output demand is explicit; time bucket prevents a closed admission count.admitted count = Unavailable if any binding request/token/time bucket is missing
Boundary: One model’s cap cannot be generalized to every catalog row.		UNAVAILABLE — time bucket missing.
16 concurrent strict extractions + retry waveconcurrency=16; schema=sha256:2f11; request cap=joined; retry wave=5; token cap=Unavailable; binding=unknownSchema and request identity join; token cap remains the unresolved binding constraint.binding constraint = min(request, token, time); unknown binding ⇒ Unavailable
Boundary: Retry count cannot convert an unknown token envelope into capacity.		UNAVAILABLE — token envelope unknown.

Wafer-scale claim-to-observation ledger

Frozen cerebras fixture — Claim source/date separated from matched measurements and acceptance; first-party evidence checked 2026-08-14; unresolved joins render Unavailable.

Method: every calculation uses only the joined fields shown in its row; missing joins fail closed. First-party source: Cerebras Inference documentation. Verified: 2026-08-14.

FixtureNamed inputsObservationFormula / boundaryDecision
short answer + 4K generationclaim source=Cerebras docs; claim date=2026-08-14; endpoint/model/tier=joined; n=12; TTFT=71ms; generation=1,020 tok/s; accepted=12/12Matched measurements support this fixture only; claim and observation remain separate fields.accepted coverage = 12/12 = 100%; useful time = queue + TTFT + generation
Boundary: Wafer-scale or advertised speed is not application success.		PASS — fixture-local observation.
long-context synthesis + strict schemacontext=96K; evidence shards=8; schema=sha256:7de2; queue=Unavailable; total=Unavailable; accepted=7/8; repair=1Seven artifacts pass; one requires manual repair and queue/total timing is missing.acceptance coverage = 7/8 = 87.5%; repair count remains visible
Boundary: No total-time claim is emitted from missing queue evidence.		PASS WITH REPAIR — not full acceptance.
three-tool loop + 64-request bursttool IDs=3/3; concurrency=64; accepted=52; errors=12; retries=4; sample=64; observation window=2026-08-14Tool IDs join; burst coverage is 52/64 and errors/retries remain part of the result.completion coverage = 52/64 = 81.25%; error rate = 12/64 = 18.75%
Boundary: Hardware or claim speed cannot erase application-level burst errors.		FAIL — burst decision boundary breached.

Route owner CTA: test this decision path at All AI Ask. This surface does not transfer evidence to another host, model, region, tier, artifact, or scenario.

Try Cerebras for free

Run real prompts against every current Cerebras model, and every other provider on this site, in one workspace.

Try It Free