How the numbers on this site are made

We list 46 models, including 13 that their providers have since retired; 29 are current and eligible for recommendations. Each price is the provider's own list price for its first-party API — standard tier, shortest-context band — checked against the provider's pricing page, and every model page shows the date it was last verified. We do not use reseller prices, averages or remembered figures.

Where a provider has announced a change, we show it: a promotional price is marked with the date it is guaranteed until and the regular price behind it, and an introductory price with the date it ends. DeepSeek charges half outside its weekday peak hours; we list the peak rate, because that is the one a production workload can plan on. Open-weights models such as Llama have no first-party price — hosts set their own — so their pages say so.

Providers retire models, and a retired model is not a cheaper option — requests to it fail. Every model carries a status taken from the provider's deprecation page: active, deprecated (a shutdown date has been announced), retired (no longer served), or superseded (open-weights models whose maker has moved on to a newer generation).

Retired models keep their pages, because people still search for them, but those pages say when the model was shut down and point to the closest current model. That suggestion is ours — chosen by role and price — and is not always the provider's official replacement. Only active models are ever recommended by the site's calculators.

Providers charge per token, but they do not cut text into tokens the same way. Anthropic documents that Claude 4.7 and later models produce about 30% more tokens for the same text than earlier Claude models. Comparing prices per token therefore understates what those models cost for a given piece of work.

We count text with o200k, the tokenizer of GPT-4o and later OpenAI models, and multiply by each model's tokenizer factor. Where a tokenizer is published (OpenAI, DeepSeek) or the provider offers a token-counting endpoint (Anthropic, Google, xAI), the factor is measured: we run our 25-language parallel corpus through it and compare it with o200k on the same English texts. Until a tokenizer has been measured we use an estimate — for Claude 4.7 and later, 1.43: an estimated 1.1 for the previous Claude tokenizer times Anthropic's documented 1.3. The table shows which is which. English is the basis because most API traffic is English; other languages can differ far more, and the language cost pages show each one.

The coding, reasoning, extraction, creative and vision scores (0–100) are iOPTERA's editorial composites. They place each model relative to the others using public benchmark results and providers' published evaluations, so that the calculators have something to rank on. They are not benchmark results, we do not run the evaluations ourselves, and a difference of a point or two means nothing.

When newer models raise the bar, older ones are re-scored downward rather than kept at their launch-day numbers, so scores across generations stay comparable. When we suggest a cheaper alternative, it has to score within two points of the model it replaces — a saving that requires a capability downgrade is not presented as a saving.

Unless a page says otherwise, a blended price assumes 70% input and 30% output tokens, the workload is measured in o200k tokens, and each model is charged for the tokens its own tokenizer would produce. Yearly and long-run comparisons use the price after any announced promotion or introductory rate ends.

Estimates leave out what depends on your setup: prompt caching, batch discounts, long-context surcharges above 200K–272K tokens, and the hidden reasoning tokens some models spend. Treat the figures as a comparison between models for the same work, not as a forecast of your invoice.