Artificial Analysis releases Intelligence Index v5: open models fall 12 points behind

Artificial Analysis has released Intelligence Index v5, the biggest update to its composite ranking since January. The index now has nine evaluations, and 60% of its weight sits on tests with private datasets or hidden answers. This guards against leakage into training data. Claude Opus 5.5 (max) leads with 58 points, though no model wins all nine tests.

How the index was rebuilt

v5 has two new benchmarks and three updated ones. Four categories carry 25% each. The breakdown:

• Knowledge Work (25%): AA-Briefcase (15%), AutomationBench-AA (5%), GDP.pdf (5%).
• Coding (25%): FrontierCode 2.0 (15%) and Terminal-Bench 4.0 (10%).
• Science (25%): Terminal-Bench-Science (15%) and a private subset of CritPt v2 (10%).
• General (25%): AA-Omniscience v1.1 (15%: 10% accuracy, 5% absence of hallucinations) and HLE-Diamond (10%).

FrontierCode 2.0 is a private Cognition benchmark of 132 tasks drawn from real open-source repositories. A result counts only if a maintainer would accept the agent's patch.

Failing any blocking check scores zero. Terminal-Bench-Science contains 70 research tasks in the terminal, and a task counts as solved only if all tests pass. HLE-Diamond replaces Humanity's Last Exam: it is a cleaned-up 1000-question version from CAIS and Scale AI, split evenly between reasoning and knowledge.

AA-Omniscience v1.1 was purged of saturated and public questions, with new ones added and ambiguous ones fixed. The average score dropped by about 2 points.

CritPt v2 is scored on a closed set of the 40 hardest questions out of 70 physics questions: the 30 that seven frontier models solved in all five attempts were removed.

GDPval-AA v2.1, SciCode and AA-LCR v1.1 have left the index. AA-Briefcase, AutomationBench-AA, GDP.pdf and Terminal-Bench 4.0 remain.

Who leads and who lags

Models from Anthropic and OpenAI hold the top five places. Gemini 4 Argon (high) follows with 50.

• Claude Opus 5.5 (max): 58 overall. Leads FrontierCode 2.0 (46%), HLE-Diamond (68%) and AA-Omniscience v1.1 (48). Costs $10.04 per index task.
• GPT-6 Astra (max): 57. Best on Terminal-Bench-Science (63%) and CritPt v2 (72% at high effort). $6.24 per task.
• GPT-6.1 Sol (max): 55. $1.27 per task.
• GPT-6.1 Terra (max): the cheapest of the top ten, at $0.96 per task.
• Claude Fable 5.1 (max): the most expensive, at $13.04. Scored 60% on HLE-Diamond.
• Claude Sonnet 5.5: leads AA-Briefcase (1812 Elo), with 42% on FrontierCode 2.0.
• Kimi K3.1 Pro: best on Terminal-Bench 4.0 (67%) and the best open model overall (46), but its weights have not yet been published.
• Gemini 4 Argon: leads AutomationBench-AA (78%).
• Muse Wonder Full: leads GDP.pdf (33% all-pass).
• DeepSeek V4.1 Flash 1017 and GLM-5.3 (max): 32 each, the best among open models with published weights. DeepSeek scores 28% on FrontierCode 2.0 and 13% on Terminal-Bench-Science.
• Claude Haiku 5.5 (max): 36, which is higher than any open-weights model. Scored 20% on Terminal-Bench-Science.

The open-model gap grew from 3 points in v4.3 to 12. It is widest on HLE-Diamond (Kimi K3.1 Pro 37% vs. 68%) and Terminal-Bench-Science (34% vs. 63%). Counting only models with published weights, the gap is 26 points. Most open models scored below 10% on Terminal-Bench-Science.

The price of intelligence and the appetite for tokens

GPT-6.1 Sol trails Opus 5.5 by just 3 points at a cost per task roughly eight times lower. Effort mode now matters more. Opus 5.5 scores 35 on low and 58 on max, while cost rises tenfold ($0.98 vs. $10.04). Sonnet 5.5's spread grew from 20 to 29 points, and GPT-6.1 Sol's grew from 10 to 20.

Tasks have gotten heavier. Median output tokens per task rose by two-thirds, from 35k to 59k. The agentic tests (Terminal-Bench-Science, Terminal-Bench 4.0, AA-Briefcase, FrontierCode 2.0) carry 55% of the weight but consume about 90% of the tokens. Terminal-Bench-Science alone accounts for about a third.

Artificial Analysis has separately launched Progress Index v1. It places models released since November 2022 on a single continuous scale, which makes it possible to compare models separated by years. Intelligence Index is still the one to use when choosing a model for current tasks. In developing it, the authors drew on the Epoch Capabilities Index and thanked Epoch AI and Anthropic.

The practical takeaway for teams: choosing a model by overall score is no longer enough. For code and science automation, look at FrontierCode 2.0 and Terminal-Bench-Science, and for budgeting, calculate cost per task and effort mode. Open models are still suitable for simple scenarios but fall well behind in long agentic chains. Before going to production, run your own tasks at several effort modes.