LLM Cost Calculator — Real Monthly Cost, Not Just Rate

List price lied to you

These models move the most once verbosity is priced in — a model that talks more costs more, regardless of its list rate.

• GLM 4.7 (Cerebras) is priced #39 by list rate but #53 once its 7.55× verbosity is billed — $1,273 list vs $2,533 effective.
• GLM-5.2 is priced #35 by list rate but #47 once its 3.87× verbosity is billed — $980 list vs $1,864 effective.
• Mistral Small 3.1 is priced #12 by list rate but #8 once its 0.85× verbosity is billed — $114 list vs $108 effective.
• Gemini 3.1 Pro is priced #48 by list rate but #44 once its 0.63× verbosity is billed — $1,800 list vs $1,489 effective.
• DeepSeek V4 Flash is priced #8 by list rate but #11 once its 2.60× verbosity is billed — $86.80 list vs $118 effective.
• GPT-OSS 120B (Cerebras) is priced #17 by list rate but #20 once its 2.33× verbosity is billed — $221 list vs $290 effective.

The formula, published

outputTokensBilled = outputTokens × (verbosityIndex ?? 1)
inputCost  = inputPerM  × inputTokens        / 1e6
outputCost = outputPerM × outputTokensBilled / 1e6
effective  = (inputCost + outputCost) × callsPerMonth

with caching: inputCost × (1 − cacheableInputPct × 0.9)
with batching: (inputCost + outputCost) × (1 − batchDiscountPct/100)

90% cache-read saving is a stated modelling assumption, not a per-provider sourced figure — providers publish cache-read discounts between 75% and 90% off. The free-form estimator above assumes 30% of input is cache-eligible; pick a workload preset below for a shape-specific assumption instead.

Coverage: 21 of 63 priced models have a measured verbosity index (2+ graded runs). The rest render with a verbosity of — and are shown at unadjusted list price.

Cost by workload shape