Where the numbers come from

Models.dev

Primary source, republished with credit. The project is anomalyco/models.dev, MIT license. The refresh reads api.json and models.json. Provider docs are linked from model pages. Benchmark scores in that dataset are not republished here.

OpenRouter

GET /api/v1/models is fetched as an internal cross-check. Its fields are not stored and not shown on pages. The committed summary records our model id and which of input, output, cache read, or cache write disagreed, without their numbers.

Last check compared 188 models. 124 were within 2% or $0.02 per 1M tokens.64 disagreed. 298 of our priced models were absent. Disagreements do not replace the Models.dev price.

OpenRouter's terms have not been reviewed. Until they are, nothing from that API is published.

Not used yet

• Provider pricing pages (OpenAI, Anthropic, Google, and others). Their terms are unchecked, so they are not scraped.
• The Hugging Face Hub API. Weight links that Models.dev already publishes are kept; download counts are not pulled.
• Artificial Analysis and LLM Stats. Both restrict republishing.