About

Information for comparing AI models is spread across launch posts, model cards, papers, pricing pages and leaderboards, each with its own format and date. ModelBench puts benchmark scores, prices, context windows, speed and release history side by side, and links every number back to where it came from.

It is meant for shortlisting: narrowing a large field to a few candidates worth testing on your own task. It is not a substitute for that test.

Scores and per-benchmark results come from published, citable sources; prices from public model listings and the labs' own pricing pages; human-preference ratings from public arena datasets; open-weights metadata from model repositories. The full list, with what each source is used for and its licence, is on Sources & credits.

Overall and category scores are reproduced as published. ModelBench does not re-weight them or add its own judgement.

• Speed measurements, coming soon. Time to first token and output speed will come from our own timed runs, never copied from other sites. Until then speed columns show a dash.
• Prices checked against the labs. Prices confirmed on a lab's own pricing page are marked Verified.
• A release record with sources. Each release date names the page it came from, and dates confirmed by the provider are marked as such.
• Open data. Everything shown is downloadable as JSON from Data & API.

• Score ranges. Each score has a 90% interval. Models whose intervals overlap are not reliably different, whatever their rank.
• Verified price. The price matches the lab's own pricing page. Other prices come from public listings and can differ by provider.
• Provider-confirmed date. The release date is backed by the provider's own announcement, documentation or changelog.
• Missing values stay missing. A model without a published result shows a dash. Nothing is estimated to fill the gap.

The data is rebuilt automatically every 12 hours: new models, prices, ratings and release records are pulled from each source, checked and published. Every page shows the data as of the last successful build, currently Oct 20, 2026.

The release record currently covers 659 releases from 114 providers.

• Negotiated, enterprise, regional or volume pricing. Only public list prices are shown.
• Private or unpublished evaluations, and results whose publishers do not allow redistribution.
• How a model performs on your data. Rankings describe public benchmarks, not your workload.

ModelBench is non-commercial. It carries no advertising, accepts no sponsorship and offers no paid placement. Rankings follow the published data, not the preferences of a model provider, a sponsor or the maintainer.

Articles on the blog state the date their data was taken, link factual claims to a primary source where one exists, and separate what a provider has announced from what we have observed.

Found a wrong number, a missing model or a broken link? Send a message through any of the profiles linked at the bottom of every page, with the page and, if possible, the source that shows the correct value. Corrections are made in the data, so they appear everywhere on the next build.