Artificial Analysis Intelligence Index v5 explained: what changed, what is private, and what a point of score costs

Artificial Analysis replaced most of its Intelligence Index with nine evaluations on October 27, 2026, in version v5, with about 60% of the tests private or held out, and Claude Opus 5.5 (max) leads the new index at 58, ahead of GPT-6 Astra (max) at 57, GPT-6.1 Sol (max) at 55 and Gemini 4 Argon (high) at 50. By Artificial Analysis's own cost figures, GPT-6.1 Sol at max costs about 2 cents per index point while Claude Opus 5.5 at max costs about 17 cents, roughly 7.5 times as much, and effort settings swing Opus 5.5 from a score of 35 at low effort ($0.98 a task) to 58 at max ($10.04 a task). The best downloadable open-weights model scores 32, a 26-point gap to the leader, while Kimi K3.1 Pro scores 46 but has not released its weights.

Artificial Analysis Intelligence Index v5 explained: what changed, what is private, and what a point of score costs On October 27, 2026, Artificial Analysis replaced most of its Intelligence Index with nine evaluations, about 60% of them private or held out. This page lays out the new tests and weights from its methodology page and turns its own cost figures into a price per point of score, which is the number most teams actually pick models on. This explains reporting by Artificial Analysis, "Artificial Analysis Intelligence Index v5" October 27, 2026 https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5 . Read the original first: https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5 https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5 In one minute - v5 swaps in FrontierCode 2.0 and Terminal-Bench-Science, upgrades HLE to HLE-Diamond, and drops GDPval-AA, SciCode and AA-LCR. Four categories, 25% each. - Top scores: Claude Opus 5.5 max 58, GPT-6 Astra max 57, GPT-6.1 Sol max 55, Gemini 4 Argon high 50. - By AA's own cost figures, Sol at max costs about 2 cents per index point; Opus 5.5 at max costs about 17 cents, roughly 7.5 times as much. - Effort settings move scores a lot: Opus 5.5 scores 35 at low effort and 58 at max, for $0.98 and $10.04 a task. - The best open weights you can download today score 32, a 26-point gap to the leader. Kimi K3.1 Pro scores 46, but its weights are not out yet. What changed in v5 Artificial Analysis calls v5 its biggest update since v4 in January. The index now has four categories at 25% each: Knowledge Work, Coding, Science and General. - Added: FrontierCode 2.0, Cognition's private set of 132 tasks in real open-source repos, graded on whether a maintainer would merge the change, and Terminal-Bench-Science, 70 agentic research tasks in a terminal. - Upgraded: HLE-Diamond, a 1,000-question version of Humanity's Last Exam; AA-Omniscience v1.1; and CritPt v2, a private hard subset of 40 physics questions. - Removed: GDPval-AA v2.1, SciCode and AA-LCR v1.1. The announcement gives no reason. It also launched a separate Progress Index v1 for comparing models across years. AA says the Intelligence Index stays the measure to use for picking a model today. The nine tests and their weights This table comes from AA's methodology page. The last column is what that page says about each test set; where it says nothing, we say so rather than guess. v5 evaluation weight tasks repeats test set per AA's pages AA-Briefcase v1.1 15% 91 1 not stated AutomationBench-AA 5% 657 1 private held-out split GDP.pdf 5% 100 5 not stated FrontierCode 2.0 15% 132 3 Cognition's private benchmark Terminal-Bench 4.0 10% 66 3 not stated Terminal-Bench-Science 15% 70 3 not stated CritPt v2 hard subset 10% 40 5 private list of 40 of 70 AA-Omniscience v1.1 15% 3,000 1 public subset on Hugging Face HLE-Diamond 10% 1,000 1 not stated The announcement says about 60% of the index now runs on private test sets or held-out answers. The methodology page names private sets for AutomationBench-AA and CritPt v2, Cognition calls FrontierCode private, and AA-Omniscience has a public subset. It does not label the rest one by one, so we cannot rebuild the 60% from these pages. Private sets cut one risk, that a model was trained on the answers. They add another: nobody outside AA can rerun the numbers. What a point of score costs AA publishes the cost to run one index task for several top models. Divide by the score and you get a price per point: Cost per Intelligence Index task, from Artificial Analysis October 27, 2026 Model and setting Index $ per task $ per index point Claude Opus 5.5 max 58 10.04 0.173 GPT-6 Astra max 57 6.24 0.109 GPT-6.1 Sol max 55 1.27 0.023 Claude Opus 5.5 low 35 0.98 0.028 $ per index point = $ per task / index score our division, not an AA metric Two things stand out. GPT-6.1 Sol at max is within 3 points of Opus 5.5 at max, at about an eighth of the cost per task, which AA also notes. And Opus 5.5 at low effort is not the bargain it looks like: it costs a little less per task than Sol at max, $0.98 against $1.27, but scores 20 points lower, so each point costs more. Effort settings now matter as much as model choice. AA says the gap between a model's lowest and highest effort grew in v5: from 20 to 29 points for Claude Sonnet 5.5, and from 10 to 20 for GPT-6.1 Sol. Among the 10 highest-scoring models, AA lists GPT-6.1 Terra max as cheapest at $0.96 a task and Claude Fable 5.1 max as the most expensive at $13.04. It does not give Terra's index score in the announcement, so we left it out of the table. Where the tokens go Every model uses more output tokens per task than it did under v4.3. AA puts the median at about 59,000, up from 35,000. Four agentic tests, Terminal-Bench-Science, Terminal-Bench 4.0, AA-Briefcase and FrontierCode 2.0, carry 55% of the weight but a median of about 90% of the output tokens. Terminal-Bench-Science alone is about a third. So the index now mostly measures, and mostly pays for, long agent runs. If your workload is short answers, the per-task costs above will overstate your bill. Open weights fell further behind AA says the gap between the best open weights model and the leader widened from 3 points in v4.3 to 12 points in v5. That 12 uses Kimi K3.1 Pro at 46, whose weights are not released yet. Among weights you can download today, DeepSeek V4.1 Flash 1017 and GLM-5.3 lead at 32, below Claude Haiku 5.5 at 36. The gap is widest on HLE-Diamond and Terminal-Bench-Science. Who is affected | Case | Status | |---|---| | Teams that picked a model on v4 scores | Rankings were recomputed on new tests. Recheck before your next model decision. | | Budgets based on max effort | AA's per-task costs vary more than 13 times across the top 10. Effort setting changes both score and cost. | | Open weights users | Best released weights score 32 against a leader at 58. | What to do - Before switching models, compare dollars per point at the effort setting you would really run, not the max score. - Run your own short eval on your own tasks. Index tasks are long agent runs that may not look like your workload. - Check the methodology page for which tests are private before you treat a score as independently checkable. What is still unknown - Which of the nine tests make up the 60% AA calls private or held out. The pages we read label only some of them. - Why GDPval-AA, SciCode and AA-LCR were dropped. The announcement does not say. - Index scores for every model in the cost list, such as GPT-6.1 Terra. Without them we could not compute their price per point. - Dollars per point is our division of AA's published numbers, not a metric AA publishes. Sources - Artificial Analysis, "Artificial Analysis Intelligence Index v5" October 27, 2026 https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5 — the original report - Artificial Analysis, Intelligence Index methodology v5.0 https://artificialanalysis.ai/methodology/intelligence-benchmarking AI News Report https://theainewsreport.com/ · every headline, every morning.