Claude Opus 5.5 Tops Artificial Analysis Intelligence Index v5, Now Mostly Built on Private Tests
The overhauled index puts 60% of its weight on private or held-out evaluations, adds new coding and science benchmarks, and puts Anthropic one point ahead of OpenAI.
Published
Last updated / Fact-checked on
Update, Oct. 28: Google's Gemini 4 Carbon has since taken the No. 1 spot on Intelligence Index v5 with a score of 59, one point ahead of Claude Opus 5.5 (max). Read our story on Gemini 4 Carbon.
Claude Opus 5.5 (max) is the top model on Artificial Analysis's new Intelligence Index v5, scoring 58. OpenAI's GPT-6 Astra (max) is one point behind at 57, and GPT-6.1 Sol (max) is third at 55. The independent benchmarking firm published the update on Oct. 27. It calls v5 its biggest upgrade since v4 launched in January.
The headline change is how the scores are made. According to Artificial Analysis, 60% of the index now comes from evaluations with private test sets or held-out answers. That is the first time most of the index has been built on private data.
Why private tests matter
Public benchmarks have a contamination problem. Once test questions are online, they can end up in a model's training data, so a high score may show memorization more than skill. Keeping tasks private makes that harder.
There is a trade-off. Outside researchers can't inspect private questions, so readers have to trust the benchmarker's grading more. Private tests reduce contamination, but they don't remove the need for independent checks.
What's in v5
Artificial Analysis added two benchmarks:
- FrontierCode 2.0, Cognition's private coding benchmark: 132 tasks in real open-source repositories, graded on whether a maintainer would merge the agent's change. Any failed blocking check scores zero.
- Terminal-Bench-Science: 70 expert-reviewed scientific research tasks run in a terminal. A task passes only if all of its tests pass.
It also upgraded HLE-Diamond (a refined 1,000-question version of Humanity's Last Exam), AA-Omniscience v1.1 and a private 40-question hard subset of CritPt v2. GDPval-AA, SciCode and AA-LCR were dropped. The nine evaluations sit in four equally weighted categories:
| Category | Weight | Evaluations |
|---|---|---|
| Knowledge work | 25% | AA-Briefcase, AutomationBench-AA, GDP.pdf |
| Coding | 25% | FrontierCode 2.0, Terminal-Bench 4.0 |
| Science | 25% | Terminal-Bench-Science, CritPt v2 hard subset |
| General | 25% | AA-Omniscience v1.1, HLE-Diamond |
The leaderboard
| Model | Index v5 score | Cost per index task |
|---|---|---|
| Claude Opus 5.5 (max) | 58 | $10.04 |
| GPT-6 Astra (max) | 57 | $6.24 |
| GPT-6.1 Sol (max) | 55 | $1.27 |
| Gemini 4 Argon (high) | 50 | — |
Anthropic and OpenAI hold the top five spots. No model leads every test. Opus 5.5 leads three evaluations: FrontierCode 2.0 (46%), HLE-Diamond (68%) and AA-Omniscience v1.1. GPT-6 Astra leads both science tests, with 63% on Terminal-Bench-Science and 72% on the CritPt v2 hard subset.
Open-weights models fall further behind. The best, Kimi K3.1 Pro (weights not yet released), scores 46. That's 12 points behind the leader, up from a 3-point gap in v4.3.
What it means for developers
The cost column matters as much as the scores. By Artificial Analysis's numbers, GPT-6.1 Sol scores three points below Opus 5.5 at about an eighth of the cost per task. For teams running agents at high volume, that gap can matter more than a small difference in score.
Effort settings matter too. Opus 5.5 scores 35 at low effort and 58 at max, and max costs about 10 times as much per task ($0.98 vs. $10.04). A model's ranking is really a ranking of one model at one setting.
A practical way to use the index:
- Shortlist two or three models near the top that fit your budget.
- Check the category that matches your work. For coding agents, look at FrontierCode 2.0. For research tools, look at the science tests.
- Test on your own tasks at several effort levels, and track success rate, latency and total cost.
The limits
v5 scores aren't comparable with earlier versions, because the tests and weights changed. To track progress over time, Artificial Analysis launched a separate Progress Index v1, which places models released since November 2022 on one continuous scale. It says the Intelligence Index stays its main measure of today's models.
Reactions collected by Techmeme were split. Some researchers said v5 matches their own experience better, while some open-model users disputed the lower scores. A single leaderboard is a useful starting point, not a replacement for testing on your own work.
How we reported this
We read Artificial Analysis's full announcement and checked every score, weight and cost figure against it. We cross-checked the summary on Techmeme. DailyMech has not rerun the benchmarks. The developer checklist is our own analysis.
Sources
- Artificial Analysis, "Announcing Artificial Analysis Intelligence Index v5" (Oct. 27, 2026): https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5
- Artificial Analysis, Intelligence benchmarking methodology: https://artificialanalysis.ai/methodology/intelligence-benchmarking
- Techmeme, Intelligence Index v5 coverage (Oct. 28, 2026): https://www.techmeme.com/261028/p27
DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.