Models
Each model is versioned with a parsing recipe containing the model's prompt and instructions. ParseBench is used to benchmark each model's capability and pricing estimates.
Pass a model's ID, like anthropic/claude-opus-5-5, as model in your parse request.
Overall is the mean of the five categories. See parsebench.ai for how each category is scored and the documents used.
Per 1K pages is what 1,000 typical pages cost at these prices, from the tokens each model used per page across ParseBench's documents. Your pages may use more or fewer tokens.
Avg per 1K pages is what 1,000 typical pages cost at these prices, measured across ParseBench's documents. Your pages may use more or fewer tokens.
Typical page is the median time one page takes, measured on real requests. See Speed below.
Up to 50 pages per sync request, and 500 per async request. Prices as of version 2026-10-06.
Token prices
Pages are charged by the tokens they use.
Most per page is the highest a page will be charged. Every parse request requires your account balance to cover the maximum charge, with the true charge settling after. Failed pages are never charged.
Speed
How long one page takes, measured on every successful page each model's current version has parsed.
Models with a cold start run on GPUs we start when needed. After one sits idle, the next request waits up to 45 seconds for it to start. If it takes longer, the request fails with model_starting and isn't charged, so retry in a minute or two.