Artificial Analysis has updated the way it evaluates language models. Its Intelligence Index v5 combines tests of knowledge work, coding, science and general reasoning, with a greater emphasis on tasks intended to reflect practical use.
What changed?
In its v5 announcement, Artificial Analysis describes nine evaluations divided equally among four areas: knowledge work, coding, science and general capabilities. It added FrontierCode 2.0 and Terminal-Bench-Science, upgraded three tests and removed others. The organization says 60% of the index is based on private test sets or held-out answers.
This is one organization's composite measure, not a universal definition of intelligence. Scores depend on the chosen tests, weights, model settings and evaluation methods. Results from different index versions should not be compared as if they were the same scale.
The initial rankings
In the October 28 snapshot after Google's announcement, Gemini 4 Carbon led with 59, followed by Claude Opus 5.5 at 58 and GPT-6 Astra at 57. Carbon was evaluated at high reasoning; the earlier v5 announcement lists Opus and Astra at max. Carbon was not publicly available at the time. Artificial Analysis's Carbon analysis confirms its overall lead, while noting it did not lead every individual area. Rankings can change as models are added or retested.
Among models with released weights, the v5 launch report lists DeepSeek V4.1 Flash 1017 and GLM-5.3, both at max reasoning, jointly leading at a displayed 32 points. The live leaderboard lists DeepSeek first and GLM second, but the displayed scores are equal; the ordering alone does not establish a full-point advantage. This comparison excludes models whose weights were announced but had not yet been released, such as Kimi K3.1 Pro at the time of the report. “Open weights” describes weight availability; individual model licences can differ.
The Artificial Analysis site has the latest leaderboard and methodology. Its numbers are useful for comparison, but the best model for a particular job also depends on quality on that task, access, speed and cost.
Translated from the original Chinese article, published October 28, 2026. Rankings are attributed to the cited October 2026 evaluations.
Comments NOTHING