LLM leaderboard: benchmark scores from labs and maintainers

This LLM leaderboard lists benchmark scores for current AI models exactly as they were published: by the lab in its model announcement or model card, by the benchmark’s maintainer (such as the ARC Prize Foundation), or by an independent evaluator (such as Artificial Analysis). Each row is one score for one model on one benchmark, with the setting the source states, the publisher, the publication date and a link to the source page.

Read the rows with four caveats:

• Self-reported and maintainer-run scores are not directly comparable. A lab picks its own harness, prompts and effort level; a maintainer or independent evaluator runs every model the same way. The “Reported by” column tells you which is which.
• Settings differ. Effort level, tools, harness and context limits change results, sometimes by many points. The “Setting” column copies what the source says.
• Versions matter. Terminal-Bench 2.1 and 4.0, or OSWorld 2.0 and 2.1, are different tests. An index such as the Artificial Analysis Intelligence Index can also change between articles, so compare index values only within the same article date.
• Every row links its source. If a number is not on the publisher’s own page, it is not here.

The table is sorted by benchmark; filter by benchmark, lab or source type, or sort by score within one benchmark. Rows are re-checked against their sources three times a week. Model release dates are in AI models, new results are covered in AI benchmark news, and how we pick sources is in our editorial standards.