Claude Opus 5.5 Leads a Rebuilt AI Index at 58 as the Open-Weights Gap Widens to 12 Points

Artificial Analysis published version 5 of its Intelligence Index on Tuesday, the largest overhaul of the composite benchmark since v4 launched in January. The structural change is the headline: 60% of the index now rests on evaluations with private test sets or held-out answers, the first time a majority of the measure is not drawn from publicly available questions.

The index keeps nine evaluations in four equally weighted buckets. Knowledge work, 25%, is AA-Briefcase at 15%, AutomationBench-AA at 5% and GDP.pdf at 5%. Coding, 25%, is Cognition's FrontierCode 2.0 at 15% and Terminal-Bench 4.0 at 10%. Science, 25%, is Terminal-Bench-Science at 15% and a private hard subset of CritPt v2 at 10%. General, 25%, is AA-Omniscience v1.1 at 15% and HLE-Diamond at 10%. GDPval-AA v2.1, SciCode and AA-LCR v1.1 are out.

The new entrants are built to resist memorization. FrontierCode 2.0 comprises 132 tasks in real open-source repositories, graded on whether a maintainer would merge the agent's change rather than on tests a model can pattern-match, with every task run three times inside Artificial Analysis's agent harness, Stirrup. Terminal-Bench-Science adds 70 expert-reviewed agentic research tasks in a terminal. HLE-Diamond is a 1,000-question refinement of Humanity's Last Exam split evenly between reasoning and knowledge. CritPt v2 contributes a fixed private list of 40 of its 70 research-level physics problems; the 30 that seven frontier models solved in all five attempts at launch were dropped.

The leaderboard reads Claude Opus 5.5 (max) at 58, GPT-6 Astra (max) at 57, GPT-6.1 Sol (max) at 55 and Gemini 4 Argon (high) at 50. No single model leads every evaluation. Opus 5.5 takes FrontierCode 2.0 at 46%, HLE-Diamond at 68% and AA-Omniscience v1.1 at 48; Astra takes Terminal-Bench-Science at 63% and the CritPt hard subset at 72%; Claude Sonnet 5.5 leads AA-Briefcase at 1,812 Elo; Kimi K3.1 Pro leads Terminal-Bench 4.0 at 67%; Gemini 4 Argon leads AutomationBench-AA at 78%; Muse Wonder Full leads GDP.pdf at 33% all-pass.

The open-weights story got worse. Kimi K3.1 Pro, whose weights are not yet released, is the best open model at 46, twelve points behind the leader, against a three-point gap in v4.3. Among models with weights actually published, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) reach 32, twenty-six points back, and Claude Haiku 5.5 (max) at 36 outscores all of them. The gap is widest on HLE-Diamond, where Kimi scores 37% against Opus 5.5's 68%, and on Terminal-Bench-Science, 34% against Astra's 63%, a track where most open models benchmarked score under 10%.

Cost is now as uneven as capability. Opus 5.5 (max) costs $10.04 per Intelligence Index task, Astra (max) $6.24 and Sol (max) $1.27, within four points of the leader at roughly an eighth of the cost. Across the ten highest-scoring models, GPT-6.1 Terra (max) is cheapest at $0.96 per task and Claude Fable 5.1 (max) most expensive at $13.04, a spread of more than thirteen times.

Effort settings matter more than they used to. Opus 5.5 scores 35 at low effort and 58 at max for ten times the cost per task, $0.98 against $10.04. Sonnet 5.5 spans 29 points from low to max, up from 20 in v4.3, and the spread on GPT-6.1 Sol doubled from 10 to 20 points. The tasks also demand more work: median output tokens per index task rose by two thirds, from 35,000 to 59,000, and the four agentic evaluations carry 55% of index weight but a median of about 90% of output tokens.

Alongside v5, Artificial Analysis launched a Progress Index v1 that places models released since November 2022 on one continuous capability scale, built with input from Epoch AI and Anthropic, to separate how much models have improved from which model to deploy today. For buyers, the practical lesson is that a single composite number now hides a thirteen-fold cost spread and a 20-plus-point swing from a settings toggle. For the benchmark, the shift to private sets is an admission that public evaluations leak into training data faster than anyone can refresh them.