Artificial Analysis updates intelligence index, ranks Google, Anthropic, and OpenAI as frontier leaders

Benchmarking AI models has always felt a little like grading students who wrote the exam. The data those models trained on and the tests used to evaluate them often overlap, making scores more a measure of memorization than genuine capability. Artificial Analysis just took a swing at fixing that problem.

The research firm launched Intelligence Index v5 on October 27, 2026, introducing a methodological overhaul that shifts 60% of the composite score to private test sets rather than web-scraped data. The stated goal is to reduce benchmark contamination, a persistent problem where models appear smarter than they are because they have effectively seen the questions before.

How the rankings shook out

At launch, Anthropic’s Claude Opus 5.5 took the top spot with a score of 58. OpenAI placed two models in the index: GPT-6 Astra at 57 and GPT-6.1 Sol at 55. The gap between first and third place across those three was narrow enough to fit in a rounding error.

That changed within 24 hours. On October 28, 2026, Google’s Gemini 4 Carbon posted a score of 59, claiming the leading position by a single point over Claude Opus 5.5.

The index covers nine evaluations across categories including knowledge work, coding, science, and general capability. No single model dominated across all nine, which is the more interesting finding buried beneath the headline rankings.

Each of the nine evaluations carries 25% of the total score, meaning the weighting across categories is equal and the composite reflects a specific set of priorities chosen by Artificial Analysis.

The widening proprietary gap

The more consequential shift in v5 is not who sits at the top. It is what is happening at the bottom of the frontier tier.

In the previous version, v4.3, the performance gap between closed proprietary models and the best open-weights models was 3 points. In v5, that gap has expanded to 12 points.

The best-performing open-weights model in the new index is Moonshot AI’s Kimi K3.1 Pro, which scored 46. Against Gemini 4 Carbon’s 59, that is a 13-point deficit. Against Claude Opus 5.5’s 58, it is a 12-point gap.

For context, open-weights models are those where the model weights are publicly released, allowing developers to run and modify them independently without relying on an API.

Why the methodology change matters

The shift to private evaluations for 60% of the composite score is the design choice that will attract the most scrutiny, for better and worse.

Better, because contamination has been a real and documented problem in AI benchmarking. When training corpora include the internet, and popular benchmarks live on the internet, there is a reasonable chance that a model’s strong performance reflects prior exposure rather than generalizable reasoning. Private test sets that models cannot have trained on are a more honest measure of capability.

Worse, because private evaluations introduce a different trust problem. Researchers and developers reading a leaderboard built on private tests cannot verify the methodology themselves. They are trusting the benchmarking organization to have designed fair, representative tasks and to have administered them consistently across models.

Artificial Analysis also added two new benchmarks in v5, including FrontierCode 2.0, which targets coding capability.