The gap, stated precisely
Two different gaps sit inside the headline number, and they support different conclusions. The 12-point gap compares the leader with Kimi K3.1 Pro at 46, a model Artificial Analysis classifies as open weights even though its weights have not been released. A reader who wants to run an open model today cannot download the entry that produces the smaller gap.
The downloadable frontier is lower. DeepSeek V4.1 Flash 1017 and GLM-5.3, the best-scoring models with released weights, sit at 32. That is 26 points behind Opus 5.5, and below Claude Haiku 5.5, Anthropic's smaller model, at 36. On the two newest evaluations the distance is wider still: DeepSeek V4.1 Flash 1017 scores 28% on FrontierCode 2.0 against Opus 5.5's 46%, and 13% on Terminal-Bench-Science against GPT-6 Astra's 63%.
The wider gap does not establish that open models became less capable. Artificial Analysis changed the benchmark composition, added more private or held-out evaluation material, and re-scored models under the revised methodology. Those changes can reduce some forms of benchmark exposure and better test performance on unfamiliar tasks, but they do not prove that every question was absent from training or isolate why the gap widened. The reported median output-token count also increased from 35,000 under v4.3 to 59,000 under v5. The appropriate conclusion is that the new index measures a different mix of capabilities, not that open models regressed.
The score is not the price
V5 is also a reminder that an index score, quoted alone, hides the two numbers a buyer actually pays. Effort settings materially change the comparison. Claude Opus 5.5 scores 35 at low effort and 58 at maximum effort, while its reported cost per index task rises from $0.98 to $10.04. The published leader comparison uses high-performing configurations, but that ranking alone does not show which model offers the best economics for a particular workload.
The second is cost per task at the top. GPT-6.1 Sol scores 55, three points behind the leader, at $1.27 per task against Opus 5.5's $10.04. For buyers considering high-volume workloads, Sol's $1.27 per index task against Opus's $10.04 is a useful comparison alongside their scores of 55 and 58. Actual value will depend on task mix, required quality, latency and pricing under the buyer's deployment conditions. Artificial Analysis publishes both score and cost information. Comparing them is more useful for procurement decisions than treating the composite ranking as a standalone measure of value.
Sources
- [1] Artificial Analysis, "Announcing Artificial Analysis Intelligence Index v5," October 27, 2026 (all figures in this article: the 60% private share, the new, upgraded and removed benchmarks, the category weights, the leader scores, the 3-to-12 open-weights gap, the released-weights scores, and the cost and effort figures; read in full this run)Read source
- [2] AI Weekly alert, "Artificial Analysis Intelligence Index v5 Widens Open-vs-Closed Model Gap to 12 Points," October 28, 2026, 08:13 UTC (corroborating summary and publication timing)Read source
- [3] ALFW, "Kimi K3.1 Pro Ranks Fourth on Artificial Analysis. Its Weights Aren't Out Yet," October 23, 2026 (the previous index's ranking, for the v4.3 comparison baseline)Read source