Artificial Analysis Intelligence Index v5
A · Strong evidence Independent benchmarking firm Artificial Analysis publishes a reweighted composite index with per-benchmark scores, costs and token use; checkable and widely cited, though the weighting choices are its own and the index runs in its own harness.
Summary based on the full text·Original: about 8 minutes to read
In short
Artificial Analysis (27 October 2026) announces version 5 of its Intelligence Index, a composite of nine evaluations weighted equally across knowledge work, coding, science and general knowledge. Most of the index now rests on private or held-out tests, which reduces the scope for models trained on the answers. Anthropic and OpenAI hold the top five places, open-weight models fall further behind, and effort settings now swing scores and cost far more than before.
Key numbers
- 60%share of the v5 index built on private test sets or held-out answers
- 58top score, Claude Opus 5.5 (max), ahead of GPT-6 Astra (57) and GPT-6.1 Sol (55)
- 3 to 12 pointsgap between the best open-weight model and the leader, v4.3 to v5
- 35 vs 58Claude Opus 5.5 score at low versus max effort, at $0.98 versus $10.04 per index task
- 35k to 59kmedian output tokens per index task, v4.3 to v5
As reported by the source.
English·Web page·opens artificialanalysis.ai
What it reports
- New: FrontierCode 2.0 (Cognition's private set of 132 real repository tasks, scored on whether a maintainer would merge the change) and Terminal-Bench-Science (70 agentic research tasks). Upgraded: HLE-Diamond (1,000 questions), AA-Omniscience v1.1, a 40-question hard subset of CritPt v2. Removed: GDPval-AA v2.1, SciCode, AA-LCR v1.1.
- Leaders: Claude Opus 5.5 (max) 58, GPT-6 Astra (max) 57, GPT-6.1 Sol (max) 55, Gemini 4 Argon (high) 50. No model leads all nine evaluations; Opus 5.5 leads three.
- Open weights: Kimi K3.1 Pro, weights not yet released, scores 46; among released weights DeepSeek V4.1 Flash 1017 and GLM-5.3 lead at 32, below Claude Haiku 5.5 (36).
- Cost per index task among the ten highest scorers ranges from $0.96 (GPT-6.1 Terra) to $13.04 (Claude Fable 5.1). GPT-6.1 Sol (max) scores within 4 points of Opus 5.5 at about an eighth of the cost ($1.27 versus $10.04).
- Effort matters more: Opus 5.5 spans 35 to 58 from low to max effort; GPT-6.1 Sol's spread doubles from 10 to 20 points.
- Agentic evaluations are 55% of the weight but a median of about 90% of output tokens.
- Also launched: a Progress Index placing models since November 2022 on one capability scale.
How they got it
Artificial Analysis runs each benchmark itself, agentic tasks in its own harness (Stirrup), FrontierCode tasks three times each. Weights and methodology are published on its site; most test sets are private.
Limits
- Weighting and benchmark selection are editorial choices; a different mix would reorder close models.
- Private test sets limit outside replication of scores.
- Results depend on the firm's harness and effort settings, not each vendor's best configuration.
- Kimi K3.1 Pro's open-weight status was prospective at publication.
Why it matters for companies
The index shows that model choice is now as much about effort setting and cost per task as about ranking: a near-top model at an eighth of the cost, or the top model at low effort scoring far lower, changes the economics of any agent workload. The widening open-weight gap matters to firms planning to self-host: on these tests, released open models sit 26 points behind the leader. Teams should benchmark at the effort level they will actually pay for.
Summary based on the full text. Read the original for the full argument and data.Read the original on artificialanalysis.ai