Artificial Analysis launches a Progress Index, and Gemini 4 Carbon leads it at 352

Artificial Analysis launched the Progress Index on October 29: one capability score for every language model it has tested since 2022, built so that models released years apart can be compared. Gemini 4 Carbon (high) leads at 351.6, and the next four sit within 3.1 points of it: Claude Sonnet 5.5 at 350.5, Claude Opus 5.5 at 350.0, GPT-6 Astra at 349.6 and GPT-6.1 Sol at 348.5.

WHY A SECOND INDEX

The Intelligence Index, now at v5.0, adds evaluations as models improve and retires the ones that stop separating leading models, so Artificial Analysis says scores from different versions cannot be compared. The Progress Index fits every model and every task on one scale with item response theory, the statistical method used to score standardized tests, and it draws on retired evaluations too. Artificial Analysis says to use the Intelligence Index to choose a model today and the Progress Index to see how capability changed over time. It names Epoch AI's Capabilities Index and a scale Anthropic published in its Claude Mythos Preview system card as prior work.

THE ORDER CHANGES

On Intelligence Index v5.0 the same five read Gemini 4 Carbon 59.4, Claude Opus 5.5 58.4, GPT-6 Astra 56.9, GPT-6.1 Sol 54.5 and Claude Sonnet 5.5 54.2. On the Progress Index, Sonnet 5.5 sits half a point above Opus 5.5.

WHAT ARTIFICIAL ANALYSIS SAYS THE CURVE SHOWS

By its own account, the top score among released models rose 35 points in 2023, 67 in 2024, 68 in 2025 and 56 so far in 2026, and the record has been raised on 11 release dates this year, about once every four weeks. It says open-weights models trail proprietary ones by about four months (median). The highest-scoring model with released weights is DeepSeek V4.1 Flash 1017 at 323.7. Kimi K3.1 Pro scores 340.7, but its weights are not out yet.

WHAT WE DO NOT KNOW

The methodology page lists no error bars, so a half-point gap at the top may not mean much. Gemini 4 Carbon still has no public API; access is limited to Google's trusted testers.