What it measures
Provider pricing on equivalent workloads — the input/output split that matters depends on your use case.
How it works
Price per million tokens published on the provider's pricing page. We track input, output, cached-input, and batch tiers. Cost-per-task estimates require multiplying these by a usage profile.
Caveats
- Token counts vary by tokeniser — the same English sentence may be 25 tokens for one model and 32 for another. Compare token-for-token, not request-for-request.
- Cached-input pricing is non-trivial and changes the effective cost dramatically for repeated prompts.
- Volume discounts and reserved capacity aren't reflected in public pricing.
- Pricing changes frequently. See our /pricing-data-policy for refresh cadence.
Primary source
Related benchmarks
- Latency benchmarks (TTFT, throughput) — Time-to-first-token and tokens-per-second — how quickly a model responds.
How we read this benchmark
Our Benchmark Methodology explains which benchmarks we cite on model pages, the caveats we apply, and why we never present any single benchmark as a verdict.