AI Model Comparison

Editorial standards

Methodology: How We Source and Date Every Figure

What we record for each model, where the prices come from, why only standard rates go in the table, and how corrections work.

Verified 2026-11-22

Every number on this site has a source and a date. This page explains what we record, how we check it, and what we deliberately leave out.

What we record for each model

  • Maker — the company that publishes the model.
  • Context window — the maximum input the API accepts, in tokens, as stated by the provider.
  • Input price and output price — USD per one million tokens, at the standard (non-batch) rate.
  • Open weights — whether the weights can be downloaded and self-hosted.
  • Date read — the day the figures were read, printed on the page that shows them.

Where the figures come from

  1. The provider's own model or pricing page is the primary source. If OpenAI, Google, Anthropic, Mistral, Thinking Machines or RadixArk publish a price, that price is what we record.
  2. Where a provider's page is not reachable from our region or does not render without JavaScript, we read the same provider data from a public model catalogue that republishes it, and we say so on the page.
  3. Public announcements — a launch post with pricing in it — are used only to confirm a figure we already have, never as the sole source for a price.

The rules we hold ourselves to

Standard rates only, in the main table. Batch, cached-input, regional, committed-use and enterprise rates are real and often much cheaper, but mixing them with standard rates makes the table incomparable. When they matter, they go in the notes under the table.

A date on every figure. "$8.00" without a date is a rumour. Every table on this site states when it was read.

No cross-benchmark comparisons. A score on one benchmark and a score on another are not on the same scale. We do not average them, rank them or put them in the same column.

Measurement and judgement are labelled differently. "$0.30 per million output tokens" is measured. "Best for coding" is a judgement based on price, context and what the provider positions the model for. When a sentence is a judgement, it says so.

No paid placement. Providers cannot pay to be included, to be ranked higher or to have a figure changed.

What we leave out, and why

Speed and latency. Tokens-per-second figures vary by region, by load and by time of day. A single number would be misleading.

Quality scores. They change with every model release and every benchmark revision, and a stale score is worse than no score. The benchmark guide explains what the tests measure and how to read them from the providers directly.

Uptime and reliability. We do not have access to provider status data and will not guess.

How often it is updated

Whenever a provider changes a listed price or ships a model that belongs in the table. Each page shows the date its figures were read; if that date is old, treat the figures as old.

Corrections

If a figure here disagrees with the provider's current listing, send the model, the figure and the date you checked to contact@aimodelcomparison.net. We verify against the primary source and correct the page and its date. Corrections are not sponsored, paid or negotiated.