Image: Artificial Analysis

Artificial Analysis has released version 5 of its Intelligence Index, the composite score many developers use to compare frontier AI models. The independent benchmarking firm calls it the biggest upgrade since v4 arrived in January, and the new results put Anthropic's Claude Opus 5.5 at the top while showing open weights models falling further behind.

Announcing Artificial Analysis Intelligence Index v5. Index v5 features 2 new benchmarks and 3 upgrades. For the first time, the majority of the Index is built on private datasets

— Artificial Analysis (@ArtificialAnlys) October 28, 2026

What changed in the test

The most important shift is where the questions come from. According to the announcement, 60% of the Index now relies on private test sets or held out answers, the first time most of it is built that way. Private sets matter because public benchmark questions can leak into training data, which lets models score well by memory rather than skill.

Two benchmarks are new. FrontierCode 2.0 is a private Cognition benchmark of 132 tasks in real open source repositories, graded on whether a maintainer would merge the change. Terminal-Bench-Science asks agents to complete 70 scientific research tasks in a terminal. HLE-Diamond, a refined 1,000 question version of Humanity's Last Exam, replaces the original, while AA-Omniscience and CritPt were also upgraded. GDPval-AA, SciCode and AA-LCR were dropped.

The Index now splits evenly across four areas. Knowledge work, coding, science and general knowledge each count for 25% of the final score.

Who leads

Claude Opus 5.5 at max effort scores 58, just ahead of OpenAI's GPT-6 Astra at 57 and GPT-6.1 Sol at 55. Anthropic and OpenAI hold the top five spots, with Google's Gemini 4 Argon next at 50. No single model wins all nine evaluations. Opus 5.5 leads three of them, including FrontierCode 2.0 at 46% and HLE-Diamond at 68%, while GPT-6 Astra tops Terminal-Bench-Science at 63%.

Open weights models lost ground. The gap between the best open model and the leader grew from 3 points under the previous version to 12. Kimi K3.1 Pro leads that group at 46, though its weights are not yet released. Among models whose weights are actually available, DeepSeek V4.1 Flash 1017 and GLM-5.3 score 32, below even Anthropic's smaller Claude Haiku 5.5 at 36.

Image: Artificial Analysis

Why cost matters

The ranking alone hides a large price gap. Among the 10 highest scoring models, cost per Index task ranges from $0.96 for GPT-6.1 Terra to $13.04 for Claude Fable 5.1. Opus 5.5 costs $10.04 per task, while GPT-6.1 Sol lands within 4 points of it for $1.27, about an eighth of the price.

Effort settings also change results sharply. Opus 5.5 scores 35 at low effort and 58 at max, with cost rising roughly tenfold. Models are also working longer, as median output per task rose from 35,000 to 59,000 tokens.

What it means for readers

For teams choosing a model, the takeaway is that the top score is not automatically the best buy. A model a few points lower can be far cheaper, and the effort setting can matter as much as the model name. Artificial Analysis also launched a Progress Index that tracks models since November 2022 on one scale. Full scores and the methodology are available on its site at the time of this writing.