Skip to content
Neural Daily

ResearchNews

AA Intelligence Index v5: Opus 5.5 leads, open gap widens

🇭🇺 Olvasd el magyarul →

Artificial Analysis released Intelligence Index v5, with 60% of the index now built on private test sets. Claude Opus 5.5 leads with 58 points, and the open weights gap grew from 3 to 12 points.

By: Mikola Máté Patrik · · 2 min read

Artificial Analysis Intelligence Index v5 announcement cover on a purple background
Image: Artificial Analysis · Press kit · Official cover image from the Artificial Analysis announcement · Source

What changed?

Artificial Analysis released Intelligence Index v5 on Tuesday (Artificial Analysis), its biggest update since v4 launched in January. For the first time, most of the composite index runs on private tests or held-out answers: 60% of the weight. That protects against contamination, making it harder to train a model to the test.

  • New: FrontierCode 2.0, Cognition's private coding benchmark. 132 tasks in real open-source repos, graded on whether a maintainer would merge the change.
  • New: Terminal-Bench-Science, 70 agentic scientific research tasks in a terminal.
  • Upgraded: HLE-Diamond (a 1,000-question version of Humanity's Last Exam), AA-Omniscience v1.1, CritPt v2 hard subset.
  • Removed: GDPval-AA, SciCode, AA-LCR.

The four categories (knowledge work, coding, science, general) get 25% each. Alongside the index, Artificial Analysis launched Progress Index v1, which tracks model capability over time on one continuous scale.

The results

  • Claude Opus 5.5 (max): 58 points, first place.
  • GPT-6 Astra (max): 57, GPT-6.1 Sol (max): 55. Anthropic and OpenAI hold the top five spots.
  • Gemini 4 Argon (high): 50.
  • Best open weights model: Kimi K3.1 Pro at 46, but its weights are not out yet. Among released weights, DeepSeek V4.1 Flash 1017 and GLM-5.3 lead at 32, below Claude Haiku 5.5 (36).

On FrontierCode 2.0, Opus 5.5 leads at 46%, with GPT-6 Astra, GPT-6.1 Terra and Claude Sonnet 5.5 all at 42%. On Terminal-Bench-Science, GPT-6 Astra leads at 63%, while the best released open weights model scores just 13%.

FrontierCode 2.0 benchmark cover image with connected lines and the number 2.0
Cognition · Press kit · Official cover image from the Cognition FrontierCode 2.0 blog post · Source

Source: Cognition. Cover image of the private FrontierCode 2.0 coding benchmark.

Why it matters

If you pick a model for agentic coding or research work, v5 is closer to real use than public benchmarks because it is harder to optimize for. The 12-point open weights gap shows up mostly on agentic tasks, so run your own tests before betting on a self-hosted agent. Old v4 scores are not comparable with the new ones.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles