AI Model Benchmarks

Read the Agentic column, not Intelligence. Intelligence is a broad quality index covering reasoning and knowledge, while Agentic aggregates coding and tool-use benchmarks such as SWE-bench Verified, HumanEval and Terminal-Bench - the tasks that actually resemble programming work. The two columns disagree often enough that the overall leader is regularly not the best coding model, so rank by Agentic and use the leaderboard above to compare.

Agentic measures whether a model can carry out multi-step work on its own: planning an approach, calling tools, writing and running code, reading the result and correcting itself. It is scored on benchmarks that give the model a real task and check whether it finished, rather than asking it a question and grading the answer. It is the best single predictor of how a model will perform inside a coding assistant or an autonomous agent.

For most everyday work, close enough that the gap rarely decides the outcome. The strongest open-weight models now land within a few points of the leading paid models on quality index, and they cost nothing to run if you host them yourself. The gap that remains is widest on agentic and long-horizon tasks. Every section on this page lists a paid and a free pick side by side so you can see the trade for your own use case.

No. Context is the ceiling on how much text a model can accept in one request, not a promise that it will use all of it well. Accuracy typically degrades toward the middle of a very long input, an effect often called lost in the middle. A model with a smaller window and good retrieval usually beats a larger window stuffed with unfiltered text, which is why what you put in the window matters more than its size.

Benchmarks are fixed test sets, and your work is not. Three things commonly explain the gap: the benchmark may overlap with the model's training data, inflating the score; benchmarks reward single correct answers while real work rewards recovering from a bad first attempt; and published scores are usually run at settings you are not using. Treat the leaderboard as a shortlist, then evaluate the top two or three on your own tasks.

They are the same measure. Intelligence is the column name used on this page for the 0-100 quality index, a single number that aggregates academic benchmarks including MMLU for broad knowledge, GPQA for graduate-level reasoning, MATH and AIME for mathematics, and ARC-AGI for abstract reasoning. It is the best available summary of general capability, which also means it hides the specific strengths the other columns expose.

The leaderboard, quick picks and use-case sections on this page are refreshed automatically every other day from published benchmark results, and the page is rebuilt each time so the numbers you see are current. Scores are best-known published values; a tilde indicates an estimated or interpolated figure. The date this page was last updated is reported in its sitemap entry.