Artificial Analysis published Intelligence Index v5 on October 27, 2026, calling it the largest upgrade since v4 in January. For the first time, 60% of the Index weight sits on evaluations with private test sets or held-out answers. On the new scale, Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra (max) at 57 and GPT-6.1 Sol (max) at 55. The open-weights gap versus the leader widens from 3 points in v4.3 to 12.
Key Facts
| Details | |
|---|---|
| Release | October 27, 2026 (Artificial Analysis article) |
| Top scores (max effort unless noted) | Claude Opus 5.5 58; GPT-6 Astra 57; GPT-6.1 Sol 55; Gemini 4 Argon (high) 50 |
| Open-weights gap | 12 points vs leader (was 3 in v4.3) |
| Top open-weights | Kimi K3.1 Pro 46 (weights not yet released); among released weights, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) at 32; Claude Haiku 5.5 (max) scores 36 |
| Private share | 60% of Index weight on private or held-out sets |
| New evals | FrontierCode 2.0 (Cognition); Terminal-Bench-Science |
| Upgraded | HLE-Diamond (replaces Humanity's Last Exam); AA-Omniscience v1.1; CritPt v2 hard subset |
| Removed | GDPval-AA v2.1; SciCode; AA-LCR v1.1 |
| Category weights | Knowledge Work / Coding / Science / General, 25% each |
| Also launched | Artificial Analysis Progress Index v1 (capability over time since Nov 2022) |
| Cost spread (top 10 cited) | GPT-6.1 Terra (max) $0.96 per Index task; Claude Fable 5.1 (max) $13.04 |
Background
v5 keeps AA-Briefcase, AutomationBench-AA, GDP.pdf, and Terminal-Bench 4.0. Coding weight is split between FrontierCode 2.0 (15%) and Terminal-Bench 4.0 (10%). Science uses Terminal-Bench-Science (15%) and a private 40-question CritPt v2 hard subset (10%). General combines AA-Omniscience v1.1 (15%) and HLE-Diamond (10%). Knowledge Work uses AA-Briefcase (15%), AutomationBench-AA (5%), and GDP.pdf (5%).
FrontierCode 2.0 covers 132 tasks in real open-source repositories and grades whether a maintainer would merge the agent's change. Artificial Analysis reports Claude Opus 5.5 (max) at 46% on that suite, with several OpenAI and Anthropic models at 42%, and DeepSeek V4.1 Flash 1017 (max) as the highest released open-weights model at 28%. On Terminal-Bench-Science, GPT-6 Astra (max) leads at 63%, with Claude Opus 5.5 (max) at 59%.
Effort settings matter more than in v4.3. The lab says Claude Opus 5.5 moves from 35 at low effort to 58 at max, at roughly 10x the cost per Index task ($0.98 vs $10.04). Median output tokens per Index task rise from about 35k in v4.3 to about 59k in v5.
Reactions and Impact
THE DECODER reports that under v5, Kimi K3.1 Pro falls five places from fourth to ninth behind eight closed models, and that Grok 4.7 loses 12 points. Nathan Lambert's Interconnects note argues that closed sets limit outside audits of bad prompts, while much recent training gain still comes from RL environments that resemble public benchmarks. Artificial Analysis itself revised AA-Omniscience after removing saturated and public-set items and fixing ambiguous questions.
What We Don't Know Yet
- How stable the open-weights gap stays once Kimi K3.1 Pro weights ship (if they do)
- Whether labs will treat private-heavy indices as the default buyer guide, or keep publishing their preferred public suites
- Independent replications of FrontierCode 2.0 and Terminal-Bench-Science outside Artificial Analysis's Stirrup harness
- Exact calendar for Progress Index methodology updates beyond the launch pages
Sources
- Announcing Artificial Analysis Intelligence Index v5 — Artificial Analysis, October 27, 2026 (primary)
- New Artificial Analysis index widens the gap between open and closed models — THE DECODER, October 28, 2026
- Artificial Analysis v5 and the benchmarker's dilemma — Nathan Lambert / Interconnects, October 28, 2026
- Latent Space AINews coverage of AA Index v5 — Latent Space, October 28, 2026 (same digest; scores match the official article)