Introducing ModelBench: a leaderboard where every number has a source

ModelBench is a free, non-commercial leaderboard for AI language models. It tracks frontier models from the big labs and open-weights models alike, and it puts four things side by side: benchmark scores, prices, context windows and speed.

Where the numbers come from

• Scores and benchmark results come from BenchLM, which compiles results published in model system cards, papers and public evaluations, and attaches a citation to each one. Their data is licensed CC BY-NC 4.0, which is why ModelBench is non-commercial.
• Prices come from BenchLM's price list and from OpenRouter's public model list. Model pages show both when they differ.
• Speed is measured by us. We stream the same prompt to each model and time the response. If we have not measured a model, we show a dash rather than borrowing someone else's number.

How to read a rank

Every overall score has a 90% interval. When two models' intervals overlap, the data cannot tell them apart, even though one is listed above the other. Treat the top few places as a group of strong options, then decide on price, context and speed.

Refreshing

The data refreshes automatically about every 12 hours. New models, price changes and ranking moves are logged on the What's new page, and everything is downloadable from Data & API.