Skip to content
audiotale.net
Sign in
NewsNews

Artificial Analysis Intelligence Index v5: Claude Opus 5.5 leads at 58, open-weights gap grows to 12

The benchmark's biggest update adds FrontierCode 2.0 and Terminal-Bench-Science, moves 60% of the index to private tests, and shows cost per task varying 13x.

audiotale.net team

· 7 min read

Written by AI from the sources linked below. How we use AI

On this page
  1. Key points
  2. What happened
  3. Background
  4. The numbers
  5. What it means
  6. What we don't know yet
  7. What to watch next
  8. Sources

Artificial Analysis, the AI benchmarking firm, released version 5 of its Intelligence Index on 27 October 2026, and the headline is a wider gap between the best closed models and open-weights ones. Claude Opus 5.5 (max effort) leads at 58, ahead of GPT-6 Astra at 57 and GPT-6.1 Sol at 55. The best open-weights model, Kimi K3.1 Pro, whose weights have not been released yet, scores 46; the best released weights, DeepSeek V4.1 Flash 1017 and GLM-5.3, score 32. In the previous version of the index, the gap between the leader and the best open-weights model was 3 points; Artificial Analysis says it is now 12.

Key points

  • Artificial Analysis Intelligence Index v5 launched on 27 October 2026 with two new benchmarks, three upgraded ones, and 60% of the index weight now on evaluations with private test sets or held-out answers.
  • Claude Opus 5.5 (max) scores 58, GPT-6 Astra (max) 57, GPT-6.1 Sol (max) 55 and Gemini 4 Argon (high) 50; Anthropic and OpenAI hold the top five places.
  • The open-weights gap widens from 3 points to 12: Kimi K3.1 Pro (weights not yet released) scores 46, and the best released weights score 32, below Claude Haiku 5.5 (max) at 36.
  • Cost per Intelligence Index task among the ten highest-scoring models runs from $0.96 for GPT-6.1 Terra (max) to $13.04 for Claude Fable 5.1 (max), a spread of more than 13 times.
  • Artificial Analysis also launched a Progress Index v1 that puts models released since November 2022 on one capability scale.

What happened

Artificial Analysis describes v5 as its biggest upgrade since v4 launched in January. The index now combines nine evaluations in four groups of equal weight (25% each):

  • Knowledge Work: AA-Briefcase (15%), AutomationBench-AA (5%) and GDP.pdf (5%).
  • Coding: FrontierCode 2.0 (15%) and Terminal-Bench 4.0 (10%).
  • Science: Terminal-Bench-Science (15%) and a private 40-question hard subset of CritPt v2 (10%).
  • General: AA-Omniscience v1.1 (15%, split 10% accuracy and 5% non-hallucination) and HLE-Diamond (10%).

Two benchmarks are new. FrontierCode 2.0 is the latest version of Cognition's private coding benchmark, launched the same day, with 132 tasks drawn from real open-source repositories; an agent's change is graded on whether a repository maintainer would merge it, and any failed blocking check scores zero. Terminal-Bench-Science has 70 agentic scientific research tasks run in a terminal; a task passes only if every one of its tests passes.

Three are upgraded. HLE-Diamond replaces Humanity's Last Exam with a refined 1,000-question version from the Center for AI Safety and Scale AI, split evenly between reasoning and knowledge questions. AA-Omniscience v1.1 removes saturated and public questions, adds new ones and fixes ambiguous items; models score on average two points lower than on v1. CritPt v2 is now a private hard subset: the 70 research-level physics questions minus the 30 that seven frontier models solved in all five attempts at launch.

Removed from the index are GDPval-AA v2.1, SciCode and AA-LCR v1.1. AA-Briefcase, AutomationBench-AA, GDP.pdf and Terminal-Bench 4.0 stay.

Background

The index is a composite built by Artificial Analysis, an independent benchmarking company, and re-versioned as models saturate older tests. When Claude Opus 5.5 launched in September, OfficeChai reported that it topped the then-current version of the index at 58, five points ahead of GPT-6 Astra and Claude Fable 5.1, which were tied at 53. Under v5 Opus 5.5 still scores 58 but the field has compressed at the top: GPT-6 Astra is now one point behind, at 57.

The coding benchmark has its own history. Cognition introduced FrontierCode on 8 June 2026, with 150 private tasks written by maintainers of 36 repositories, to test whether coding agents write code good enough to merge, not just code that passes tests, and Artificial Analysis added it to its coverage with a post titled "Announcing FrontierCode on Artificial Analysis". Searches for FrontierCode 2.0 turned up only Artificial Analysis's own description, so the 132-task figure and the launch date rest on that one source.

Open-weights models have also moved around in recent weeks. We reported on Kimi K3.1 Pro's launch at 55 on the earlier version of the index, with weights due by 4 November. Under v5 the same model scores 46, which shows how much a benchmark revision, not a model change, can move a headline number. Artificial Analysis itself says Kimi K3.1 Pro's weights are not yet released.

The numbers

Artificial Analysis Intelligence Index v5, selected modelsindex points
Artificial Analysis Intelligence Index v5, selected models (index points)
index points
Claude Opus 5.5 (max)58
GPT-6 Astra (max)57
GPT-6.1 Sol (max)55
Gemini 4 Argon (high)50
Kimi K3.1 Pro (weights not released)46
Claude Haiku 5.5 (max)36
DeepSeek V4.1 Flash 1017 (max)32
GLM-5.3 (max)32

Source: Artificial Analysis, Intelligence Index v5 announcement (27 October 2026)

Cost is where the leaderboard gets interesting. Artificial Analysis reports the cost of running one Intelligence Index task, and the figures differ far more than the scores do.

Cost per Intelligence Index task, selected models (US$)US$ per task
Cost per Intelligence Index task, selected models (US$) (US$ per task)
US$ per task
Claude Fable 5.1 (max)13.04
Claude Opus 5.5 (max)10.04
GPT-6 Astra (max)6.24
GPT-6.1 Sol (max)1.27
GPT-6.1 Terra (max)0.96

Source: Artificial Analysis, Intelligence Index v5 announcement (27 October 2026)

GPT-6.1 Sol (max) lands within 4 points of Claude Opus 5.5 (max) at roughly an eighth of the cost per task, by Artificial Analysis's numbers. Effort settings matter more than before: Opus 5.5 scores 35 at low effort and 58 at max, at about ten times the cost per task ($0.98 against $10.04). Claude Sonnet 5.5 spans 29 points from low to max effort, up from 20 in v4.3, and GPT-6.1 Sol's spread doubles from 10 to 20 points.

The tasks are also heavier. Artificial Analysis says every model it measured writes more output tokens per task than under v4.3, with the median up by two thirds from 35,000 to 59,000. Four agentic evaluations (Terminal-Bench-Science, Terminal-Bench 4.0, AA-Briefcase and FrontierCode 2.0) make up 55% of the weight but a median of about 90% of the output tokens, and Terminal-Bench-Science alone accounts for about a third of the output per task.

What it means

For developers choosing a model. No model wins everything. Of the nine evaluations, Artificial Analysis says Claude Opus 5.5 leads three (FrontierCode 2.0 at 46%, HLE-Diamond at 68%, AA-Omniscience v1.1 at 48), and five other models lead the remaining six: GPT-6 Astra on Terminal-Bench-Science (63%) and the CritPt v2 hard subset (72%), Claude Sonnet 5.5 on AA-Briefcase (1812 Elo), Kimi K3.1 Pro on Terminal-Bench 4.0 (67%), Gemini 4 Argon on AutomationBench-AA (78%) and Muse Wonder Full on GDP.pdf (33% all-pass). A single index number hides that, and it hides cost.

For coding work. On FrontierCode 2.0, Opus 5.5 (max) leads at 46%, followed by GPT-6 Astra (max), GPT-6.1 Terra (max) and Claude Sonnet 5.5 (max) at 42%. The best open-weights model, DeepSeek V4.1 Flash 1017 (max), scores 28%, 18 points behind. A score of 46% on a test where the bar is "would a maintainer merge this" is a reminder that most agent output still fails review.

For open-weights models. The 12-point gap is the biggest change in the update. The widest differences are on HLE-Diamond and Terminal-Bench-Science, where Kimi K3.1 Pro scores 37% and 34% against 68% and 63% for the leaders. Among models with released weights, most scored under 10% on Terminal-Bench-Science, and the best, DeepSeek V4.1 Flash 1017, scores 13%. Part of the widening is likely the move to private test sets, which cannot be trained on; Artificial Analysis does not attribute the gap to a single cause, and neither can we.

For reading benchmarks. 60% of the index now rests on private or held-out data. That makes gaming harder, but it also means outsiders cannot re-run it: you have to trust the operator. The numbers in this article are all Artificial Analysis's own.

What we don't know yet

  • How v5 relates to the old scale. Scores under v4.3 and v5 are not directly comparable, since the evaluations changed. Artificial Analysis introduced a separate Progress Index v1 for comparing models across years; we have not examined its method.
  • Kimi K3.1 Pro's real standing among open models. Its weights are unreleased, so nobody outside the lab can verify the 46.
  • Independent replication. We found no other group that has re-run v5 or FrontierCode 2.0.
  • Which other labs' models were left out. The announcement shows a selection of the models measured, not the full table.

What to watch next

Watch the Kimi K3.1 Pro weights release, due by 4 November according to its launch announcement, and whether the open-weights score holds up when independent users can run it. Watch for the labs' responses on the private benchmarks, and for Artificial Analysis's Progress Index at artificialanalysis.ai/trends.

Sources

SaveStart the discussion

Written by

audiotale.net team

The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.

Read next