# Artificial Analysis Launches Intelligence Index v5 With Majority Private Datasets, OpenAI Models Gain, Muse Wonder, Kimi Drop > Source: > Published: 2026-10-28 04:55:27+00:00 Artificial Analysis has rolled out version 5 of its Intelligence Index, the composite benchmark that has become one of the most closely watched scorecards for frontier AI models, and this update marks the firm’s biggest overhaul since [v4 launched back in January](https://officechai.com/ai/artificial-analysis-rejigs-intelligence-index-gpt-5-2-xhigh-takes-top-spot). The headline change: for the first time, a majority of the Index, 60% of it, is now built on evaluations that use private test sets or held-out answers, a direct response to concerns about benchmark contamination as labs train on increasingly scraped web data. Alongside the Intelligence Index v5, Artificial Analysis is also launching a new Progress Index v1, meant to track model capability on a single continuous scale over time rather than resetting with each version bump. ## What’s New The v5 Index adds two fresh benchmarks and upgrades three existing ones. On the new side is FrontierCode 2.0, a private benchmark built by Cognition that tests agents on real-world coding tasks inside actual open-source repositories, graded on whether a human maintainer would actually merge the change. Tasks that fail a blocking check score zero, and the dataset stays private specifically to prevent contamination. Artificial Analysis runs every task three times using Stirrup, its own agentic harness. The second new addition is Terminal-Bench-Science, a set of 70 agentic scientific research tasks carried out entirely inside a terminal environment, with experts curating and reviewing each one. On the upgrade side, Humanity’s Last Exam has been replaced by HLE-Diamond, a refined 1,000-question version split evenly between reasoning and knowledge questions. AA-Omniscience has been bumped to v1.1, with saturated and public-set questions stripped out, fresh questions added, and ambiguous or erroneous ones revised. And CritPt has moved to v2, now scored on a private 40-question hard subset revised by the benchmark’s original authors and run with code-execution access. Four evaluations carry over unchanged: AA-Briefcase, AutomationBench-AA, GDP.pdf and Terminal-Bench 4.0. Three have been dropped entirely: GDPval-AA v2.1, SciCode and AA-LCR v1.1. ## Anthropic And OpenAI Hold The Top Five Claude Opus 5.5 (max) tops the new leaderboard with a score of 58, just ahead of GPT-6 Astra (max) at 57 and GPT-6.1 Sol (max) at 55. Between them, Anthropic and OpenAI models occupy all of the top five spots, with Gemini 4 Argon (high) the next-best from another lab at 50. It’s a continuation of the trend that’s played out across recent Index updates, where the two labs have repeatedly swapped the top spot; Anthropic’s Claude Fable 5 had [topped v4.1 with a score of 60](https://officechai.com/ai/61673/) back when it launched, and the lead has changed hands several times since, including a run when [GPT-5.5 took the top spot](https://officechai.com/ai/gpt-5-5-tops-artificial-analysis-with-score-of-60/) by breaking a three-way tie. Notably, the absolute scores across the board look lower than they did under the old v4.3.2 methodology, where Claude Opus 5.5 (max) also led but at a score of 58 against a top-20 field that was tightly bunched in the 40s and 50s. Under v5’s tougher, more private evaluation mix, that same field now spreads out more, with several previously strong performers sliding down the rankings — a sign the new private datasets are proving harder to game. ## Open Weights Models Fall Further Behind One of the starker findings in this release is how much ground open weights models have lost. The gap between the best open weights model and the overall leader has widened from 3 points to 12. Kimi K3.1 Pro, whose weights have not yet been released, is the top open weights performer at 46, but among models whose weights are actually out, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) share the lead at just 32, a full 26 points behind Claude Opus 5.5 and even behind Claude Haiku 5.5 (max) at 36. That widening gap is a notable shift from where things stood even a few updates ago, when open models were closing in on proprietary leaders rather than falling back. ## Cost Varies Wildly Among Top Models Among the ten highest-scoring models on the new Index, the cost of running a single Intelligence Index task varies by more than 13 times. GPT-6.1 Terra (max) comes in cheapest at $0.96 per task, while Claude Fable 5.1 (max) is the priciest at $13.04, underlining just how differently labs are pricing comparable levels of frontier capability. ## How The New Benchmarks Shook Out On FrontierCode 2.0, which went live the same day as this Index update, Claude Opus 5.5 (max) leads at 46%, with GPT-6 Astra (max), GPT-6.1 Terra (max) and Claude Sonnet 5.5 (max) bunched together at 42%. DeepSeek V4.1 Flash 1017 (max) is the strongest open weights entrant here too, at 28%, 18 points off the pace. Terminal-Bench-Science tells a different story, with OpenAI on top: GPT-6 Astra (max) scores 63%, followed by Claude Opus 5.5 (max) at 59% and GPT-6.1 Sol (max) at 58%. The drop-off below the leaders is steep, Kimi K3.1 Pro manages just 34%, Claude Haiku 5.5 (max) scores 20%, and most open weights models tested come in under 10%. On HLE-Diamond, Claude Opus 5.5 (max) leads at 68%, ahead of GPT-6 Astra (max) at 64% and Claude Fable 5.1 (max) at 60%. Interestingly, the leading models find knowledge questions noticeably harder than reasoning ones: Opus 5.5 and GPT-6 Astra both score 78% on reasoning, but that drops to 58% and 50% respectively on knowledge questions. AA-Omniscience v1.1, the cleaned-up version of Artificial Analysis’s hallucination and knowledge benchmark, produced scores broadly in line with the original version, with models averaging about 2 points lower (a median drop of 0.6). A new public companion dataset, AA-Omniscience-Public-v1.1, has also been released on Hugging Face. On the CritPt v2 hard subset of 40 research-level physics questions, GPT-6 Astra leads at 72% on high effort, with GPT-6.1 Sol close behind at 69% on max effort. Claude Opus 5.5 (max), despite topping the overall Index, scores 61% here. Kimi K3.1 Pro again leads the open weights field, at 45%. ## No Single Model Leads Everywhere Across the nine evaluations that make up v5, Claude Opus 5.5 leads three of them: FrontierCode 2.0, HLE-Diamond and AA-Omniscience v1.1. The remaining six are split among five other models. GPT-6 Astra tops Terminal-Bench-Science and the CritPt v2 hard subset. Claude Sonnet 5.5 leads AA-Briefcase with a 1812 Elo score. Kimi K3.1 Pro leads Terminal-Bench 4.0 at 67%. Gemini 4 Argon tops AutomationBench-AA at 78%. And Muse Wonder Full leads GDP.pdf with a 33% all-pass rate. Artificial Analysis’s own Intelligence Index has steadily evolved its composition with each release, moving away from [the GPT-5.2 (xHigh) era of v4.0](https://officechai.com/ai/artificial-analysis-rejigs-intelligence-index-gpt-5-2-xhigh-takes-top-spot) and the subsequent run of leads traded between [Claude Opus 4.6](https://officechai.com/ai/claude-opus-4-6-takes-top-spot-on-artificial-analysis-intelligence-index-has-two-point-lead-over-gpt-5-2) and OpenAI’s models, toward a benchmark suite that leans far more heavily on datasets models can’t have seen during training. With 60% of the Index now private, v5 looks designed to hold up longer before the field catches up to it.