How we rank AI models

Every ranking comes from public, citable data and is recalculated every week. Nobody can pay for a position.

Where the data comes from

Today we use the Arena Leaderboard Dataset (CC BY 4.0): millions of blind votes where people compare answers from two anonymous models and pick the better one. Each task uses the matching arena: coding votes for coding, creative-writing votes for writing, text-to-image votes for images, agent sessions for agents, search votes for web research, and document and vision votes for reading documents.

The quality score

A model's quality score is its percentile among all models rated for that task: 100 is the best rated, 0 the lowest. When a task uses several sources, we average them with fixed weights that are published with the code.

Price and value for money

Prices are the providers' list prices as published on OpenRouter. For text models we use a blended price per million tokens: three parts input to one part output, because prompts are usually longer than answers. For image models, the price per generated image. The value score combines 60% quality and 40% price, where price is ranked against the other models in the same task. Models without a known price don't appear in the value ranking.

Ties

Votes come with a margin of error. When two models' margins overlap with the leader's, we can't honestly say which is better, so we group them as tied for first.

Limited data

A model rated by fewer than 10% of the typical number of voters for that task is labelled "Limited data". It still appears in the ranking, but it can't take first place until more people have rated it.

Which models we list

Only current models: available to use today through the major model marketplace (OpenRouter) and released there in the last 12 months. Older versions are left out so the rankings stay useful.

What's coming

Speed, so you can also find the fastest model, and prices for image models.