Artificial Analysis Releases Intelligence Index v5

Artificial Analysis announced Intelligence Index v5 on October 27, adding two benchmarks and upgrading three others. The company says 60% of the index now uses private test sets or held-out answers, and calls this its biggest update since v4 launched in January. It also launched Progress Index v1, which places language models released since November 2022 on one capability scale to show change over time.

The two new evaluations are FrontierCode 2.0, a private coding benchmark with 132 tasks in real open-source repositories, and Terminal-Bench-Science, which tests 70 scientific research tasks in a terminal. FrontierCode tasks are run three times in Stirrup, an agentic harness (software for running AI agents through tasks); a task scores 0 if it fails any blocking check. Terminal-Bench-Science tasks pass only when every test passes. Three evaluations were upgraded: HLE-Diamond is a refined set of 1,000 questions, split evenly between reasoning and knowledge; AA-Omniscience v1.1 removes saturated and public questions and revises ambiguous or erroneous ones; and CritPt v2 uses a private 40-question hard subset. AA-Briefcase, AutomationBench-AA, GDP.pdf and Terminal-Bench 4.0 remain in the index, while GDPval-AA v2.1, SciCode and AA-LCR v1.1 were removed.

The index assigns 25% each to Knowledge Work, Coding, Science and General. Claude Opus 5.5 leads overall with 58 points, ahead of GPT-6 Astra at 57 and GPT-6.1 Sol at 55; Anthropic and OpenAI hold the top five places, with Gemini 4 Argon next at 50. Claude Opus 5.5 leads FrontierCode 2.0 at 46%, HLE-Diamond at 68%, and AA-Omniscience v1.1 at 48. GPT-6 Astra leads Terminal-Bench-Science at 63% and the CritPt v2 hard subset at 72%. No model leads all nine evaluations.

The top open-weights model, Kimi K3.1 Pro, scores 46, 12 points behind the overall leader; the gap was 3 points in v4.3. Its weights have not been released. Among models with released weights, DeepSeek V4.1 Flash 1017 and GLM-5.3 lead at 32, while Claude Haiku 5.5 scores 36. On HLE-Diamond and Terminal-Bench-Science, Kimi scores 37% and 34%, compared with Claude Opus 5.5 at 68% and GPT-6 Astra at 63%.

Costs per Intelligence Index task range from $0.96 for GPT-6.1 Terra among the 10 highest-scoring models shown to $13.04 for Claude Fable 5.1. Claude Opus 5.5 costs $10.04 per task, GPT-6 Astra $6.24 and GPT-6.1 Sol $1.27; GPT-6.1 Sol scores within 4 points of Claude Opus 5.5 at about one eighth the cost. Effort settings also change results: Claude Opus 5.5 scores 35 at low effort and 58 at max, while cost rises from $0.98 to $10.04. Claude Sonnet 5.5's low-to-max score spread is 29 points, up from 20 in v4.3, and GPT-6.1 Sol's spread doubles from 10 to 20.

Every measured model generates more output tokens per task than in v4.3; the median rises two thirds, from 35k to 59k. Four agentic evaluations account for 55% of index weight but about 90% of median output tokens, and Terminal-Bench-Science alone accounts for about one third of tokens per task.

Artificial Analysis announced Intelligence Index v5 on October 27, adding two benchmarks and upgrading three others. The company says 60% of the index now uses private test sets or held-out answers, and calls this its biggest update since v4 launched in January. It also launched Progress Index v1, which places language models released since November 2022 on one capability scale to show change over time.

The two new evaluations are FrontierCode 2.0, a private coding benchmark with 132 tasks in real open-source repositories, and Terminal-Bench-Science, which tests 70 scientific research tasks in a terminal. FrontierCode tasks are run three times in Stirrup, an agentic harness (software for running AI agents through tasks); a task scores 0 if it fails any blocking check. Terminal-Bench-Science tasks pass only when every test passes. Three evaluations were upgraded: HLE-Diamond is a refined set of 1,000 questions, split evenly between reasoning and knowledge; AA-Omniscience v1.1 removes saturated and public questions and revises ambiguous or erroneous ones; and CritPt v2 uses a private 40-question hard subset. AA-Briefcase, AutomationBench-AA, GDP.pdf and Terminal-Bench 4.0 remain in the index, while GDPval-AA v2.1, SciCode and AA-LCR v1.1 were removed.

The index assigns 25% each to Knowledge Work, Coding, Science and General. Claude Opus 5.5 leads overall with 58 points, ahead of GPT-6 Astra at 57 and GPT-6.1 Sol at 55; Anthropic and OpenAI hold the top five places, with Gemini 4 Argon next at 50. Claude Opus 5.5 leads FrontierCode 2.0 at 46%, HLE-Diamond at 68%, and AA-Omniscience v1.1 at 48. GPT-6 Astra leads Terminal-Bench-Science at 63% and the CritPt v2 hard subset at 72%. No model leads all nine evaluations.

The top open-weights model, Kimi K3.1 Pro, scores 46, 12 points behind the overall leader; the gap was 3 points in v4.3. Its weights have not been released. Among models with released weights, DeepSeek V4.1 Flash 1017 and GLM-5.3 lead at 32, while Claude Haiku 5.5 scores 36. On HLE-Diamond and Terminal-Bench-Science, Kimi scores 37% and 34%, compared with Claude Opus 5.5 at 68% and GPT-6 Astra at 63%.

Costs per Intelligence Index task range from $0.96 for GPT-6.1 Terra among the 10 highest-scoring models shown to $13.04 for Claude Fable 5.1. Claude Opus 5.5 costs $10.04 per task, GPT-6 Astra $6.24 and GPT-6.1 Sol $1.27; GPT-6.1 Sol scores within 4 points of Claude Opus 5.5 at about one eighth the cost. Effort settings also change results: Claude Opus 5.5 scores 35 at low effort and 58 at max, while cost rises from $0.98 to $10.04. Claude Sonnet 5.5's low-to-max score spread is 29 points, up from 20 in v4.3, and GPT-6.1 Sol's spread doubles from 10 to 20.

Every measured model generates more output tokens per task than in v4.3; the median rises two thirds, from 35k to 59k. Four agentic evaluations account for 55% of index weight but about 90% of median output tokens, and Terminal-Bench-Science alone accounts for about one third of tokens per task.