How this is scored
One benchmark at a time. On each benchmark a score is a share of the best result there, so 100 means level with the leader. Two scales need a conversion: Epoch's index is counted from its lower anchor of 130, and Arena's ratings become the expected win rate against the leader, doubled.
Weights. Each benchmark has a credibility rating out of 35: seven tests scored 1 to 5 (construct validity, scoring validity, reliability, resistance to contamination, standardisation, discriminative power and external validity). Its weight is that rating squared, then halved if the board is stale, superseded, reported only by the labs themselves, or rated Flawed by Epoch AI. The ratings are Claude's judgments, not published figures.
Labs. A lab is scored on each of the benchmarks by its best published model, and those scores are averaged with the weights. A benchmark where a lab has no result is left out of its average, which can flatter a lab that is missing from the harder ones.
Models. Each kind of work is ranked on the benchmarks that test it. A model is listed once those benchmarks carry at least % of the weight for that kind of work. "Available now" means released within months of the newest model and open to the public.
Sources and limits. Scores were read on from each benchmark's public leaderboard, mostly through mirror sites that show only a board's top 60 rows, so older and lower-ranked models can be missing. Price and speed are not counted. Compiled with Claude, an Anthropic model.
Benchmarks behind each kind of work