Methodology
How the numbers on this site are produced, and what they deliberately leave out.
Where prices come from
Every rate is taken from the provider's own public pricing page — no resellers, no aggregators, no estimates. The current sources are:
- OpenAI — https://developers.openai.com/api/docs/pricing
- Anthropic — https://platform.claude.com/docs/en/about-claude/pricing
- Google — https://ai.google.dev/gemini-api/docs/pricing
Prices are re-verified on a schedule and the verification date is stamped on every page. If a page says it was verified more than a month ago, treat it as stale and check the provider directly — we would rather tell you that than quietly show you old numbers.
How costs are calculated
Monthly cost for a workload is:
monthly = calls x ( input_tokens / 1e6 x effective_input
+ output_tokens / 1e6 x output_price )
Where a provider publishes no separate cached-input rate, cached tokens are billed at the full input rate — which is what actually happens on your invoice.
The "blended" rate shown in tables assumes 3 input tokens per output token. That ratio is a reasonable central case for production traffic, and it is wrong for your application. It exists to make a long table sortable, not to price your system. Use a calculator for that.
What this site does not do
It does not benchmark quality, latency, throughput, rate limits, or reliability. A model that is three times cheaper per token but needs two attempts per task costs more, not less, and no price table can tell you that. Cache write costs, batch discounts, fine-tuning surcharges, and image or audio token pricing are also excluded from the headline math; where they materially change a model's economics it is noted on that model's page.
Corrections
If a price here is wrong, it is worth more to us to fix it than to leave it. Send it over and it will be corrected on the next build.