Gemini 4 Carbon Takes the Lead in AI Index With Score of 59
Google's 'Gemini 4 Carbon' scored 59 points, topping an AI performance evaluation. That is 27 points higher than the best score of 32 among models with released weights. The leaderboard changed after Google's Gemini 4 Carbon appeared on the rankings on the 28th.
The 'Intelligence Index v5,' released by evaluation firm Artificial Analysis on the 27th, applies weighting across nine benchmarks and produces a single composite score from 0 to 100, with results from knowledge work, coding, science, and general domains each accounting for 25%. When the index was first published, Anthropic's Claude Opus 5.5 led with 58 points. OpenAI's GPT-6 Astra and GPT-6.1 Sol scored 57 and 55 points, respectively.
Some 60% of the v5 score is based on private test sets or evaluations whose answers are not publicly available. The rationale for raising the share of private evaluations is to reduce the impact of optimizing model performance to fit publicly known problems on the score. However, since evaluation results without public questions or answers are difficult to reproduce externally under the same conditions, and the evaluation items have changed, the score difference between v4.3 and v5 cannot be readily interpreted as a change in the models' inherent capabilities.
The gap between proprietary and open-weight models widened from 3 points in v4.3 to 12 points when v5 was first released. At the time, Kimi K3.1 Pro, which scored 46 points, was classified as the top open-weight model, but its weights have yet to be released. Among models whose weights have actually been released, the current best scores are 32 points each for DeepSeek V4.1 Flash 1017 and GLM-5.3.
Gemini 4 Carbon achieved the highest score on four of the nine benchmarks, but it was not the leader in coding or knowledge work. Since the Intelligence Index is primarily based on English-language text evaluations and does not measure actual adoption rates or developer preferences, it is difficult to conclude from the composite score alone that a particular model is best suited for all tasks.
Artificial Analysis explained on the 28th that Gemini 4 Carbon is currently available only to some users and has not yet seen a general release.
