Artificial Analysis shipped Intelligence Index v5 on October 27, 2026, and the leaderboard it produces looks different from the one you may have been using. Claude Opus 5.5 (max) leads at 58, GPT-6 Astra is one point behind, and the best open-weight model now trails the leader by 12 points instead of 3. The cause is a rebuilt benchmark: nine evaluations, four equal category weights, and roughly 60% of the score drawn from private or held-out data. This is the successor to the v4.2 update we covered in September, which had promised exactly this direction.
TL;DR #
| Question | Answer |
|---|---|
| What launched? | Intelligence Index v5, October 27, 2026, per [Artificial Analysis](https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5?ref=explainx) |
| Who leads? | Claude Opus 5.5 (max) 58, GPT-6 Astra (max) 57, GPT-6.1 Sol (max) 55, Gemini 4 Argon (high) 50 |
| What is new? | FrontierCode 2.0, Terminal-Bench-Science, HLE-Diamond, upgraded AA-Omniscience and CritPt | | What was dropped? | GDPval-AA v2.1, SciCode, AA-LCR v1.1 | | How private is it? | About 60% of weight uses private test sets or held-out answers | | Open weights? | Gap to leader widened from 3 to 12 points; best released weights score 32 | | Cost per task? | Opus 5.5 max about $10.04, GPT-6 Astra max $6.24, GPT-6.1 Sol max $1.27 |
What is in Intelligence Index v5? #
According to Artificial Analysis's methodology page, v5.0 averages nine evaluations in four categories, each worth 25%.
| Category | Evaluation | Weight |
|---|---|---|
| Knowledge Work | AA-Briefcase v1.1 | 15% |
| Knowledge Work | AutomationBench-AA | 5% |
| Knowledge Work | GDP.pdf | 5% |
| Coding | FrontierCode 2.0 | 15% |
| Coding | Terminal-Bench 4.0 | 10% |
| Science | Terminal-Bench-Science 0.1 | 15% |
| Science | CritPt v2 (hard subset) | 10% |
| General | AA-Omniscience v1.1 | 15% |
| General | HLE-Diamond | 10% |
Two evaluations are new. FrontierCode 2.0 is Cognition's private coding benchmark: 132 tasks, graded on whether a repository maintainer would plausibly merge the change. Terminal-Bench-Science has 70 agentic research tasks across five domains, run in a terminal. Three were upgraded: HLE-Diamond (1,000 questions, half reasoning and half knowledge) replaces Humanity's Last Exam, AA-Omniscience v1.1 removed saturated and public questions, and CritPt v2 became a private 40-question hard subset run with a code tool.
GDPval-AA v2.1, SciCode and AA-LCR v1.1 are gone from the index, though Artificial Analysis still lists them as additional or legacy evaluations. Our earlier posts on FrontierCode 1.1 and CritPt v2 give background on those two pieces.
The v5 leaderboard #
A podium of ranked AI models representing the Artificial Analysis Intelligence Index v5 leaderboard
These are Artificial Analysis's own numbers, at the effort setting named in brackets.
| Model | Index v5 score |
|---|---|
| Claude Opus 5.5 (max) | 58 |
| GPT-6 Astra (max) | 57 |
| GPT-6.1 Sol (max) | 55 |
| Gemini 4 Argon (high) | 50 |
| Kimi K3.1 Pro (weights not released) | 46 | | Claude Haiku 5.5 (max) | 36 | | DeepSeek V4.1 Flash 1017 (max), open weights | 32 |
| GLM-5.3 (max), open weights | 32 | Anthropic and OpenAI occupy the top five positions. For model-by-model background, see our launch coverage of Claude Opus 5.5, GPT-6.1 Sol and Gemini 4 Argon.
No model wins everything
The headline number hides a split decision. Per Artificial Analysis, Opus 5.5 leads three of nine evaluations and five other models share the rest:
- **FrontierCode 2.0:** Opus 5.5, 46%.
- **HLE-Diamond:** Opus 5.5, 68%, ahead of GPT-6 Astra (64%) and Claude Fable 5.1 (60%).
- **AA-Omniscience v1.1:** Opus 5.5, 48.
- **Terminal-Bench-Science:** GPT-6 Astra, 63%.
- CritPt v2 hard subset: GPT-6 Astra, 72% at high effort.
- **AA-Briefcase:** Claude Sonnet 5.5, 1812 Elo.
- **Terminal-Bench 4.0:** Kimi K3.1 Pro, 67%.
- **AutomationBench-AA:** Gemini 4 Argon, 78%.
- **GDP.pdf:** Muse Wonder Full, 33% all-pass.
That spread is why an aggregate should be a starting filter, not a purchase decision. Our guide to reading AI benchmarks walks through the same habit.
What does it cost per task? #
Artificial Analysis also publishes cost per task, and here the ranking inverts. Among the ten highest-scoring models it shows, GPT-6.1 Terra (max) is cheapest at $0.96 and Claude Fable 5.1 (max) most expensive at $13.04. The three leaders:
| Model | Score | Cost per task |
|---|---|---|
| Claude Opus 5.5 (max) | 58 | $10.04 |
| GPT-6 Astra (max) | 57 | $6.24 |
| GPT-6.1 Sol (max) | 55 | $1.27 |
GPT-6.1 Sol lands within four points of Opus 5.5 at roughly one-eighth of the cost, which fits the story in our Sol versus Astra cost-efficiency post.
Effort settings matter just as much. Opus 5.5 scores 35 at low effort for $0.98 per task and 58 at max for $10.04. Claude Sonnet 5.5 now spans 29 points from low to max (up from 20 in v4.3), and GPT-6.1 Sol's spread doubled from 10 to 20 points. So "which model" is only half the question; "which effort" is the other half.
Why costs went up
Artificial Analysis reports that median output tokens per task rose from 35k in v4.3 to 59k in v5. Four agentic evaluations (Terminal-Bench-Science, Terminal-Bench 4.0, AA-Briefcase, FrontierCode 2.0) carry 55% of the weight but about 90% of output tokens, and Terminal-Bench-Science alone accounts for about a third. A heavier agentic mix makes the benchmark itself pricier to run and rewards models that think and act for longer.
Why did open-weight models fall behind? #
Open-weight model distribution illustrating the widening open-weight gap on the Intelligence Index
Artificial Analysis reports the gap between the best open-weight model and the overall leader grew from 3 points in v4.3 to 12 in v5. Kimi K3.1 Pro leads open weights at 46, but its weights were not yet released at the time of the index; among models with weights you can actually download, DeepSeek V4.1 Flash 1017 and GLM-5.3 lead at 32, which is 26 points behind. Claude Haiku 5.5 at 36 outscores every released-weight model.
Trending Topics, which covered the shift, cites Kimi K3.1 Pro falling from 55 to 46, DeepSeek V4.1 Flash from 45 to 32, and Meta's Muse Spark Open from 39 to 25. The gaps are widest on the hardest tasks: Kimi K3.1 Pro scores 37% on HLE-Diamond versus Opus 5.5's 68%, and 34% against GPT-6 Astra's 63% on Terminal-Bench-Science. Yet Kimi tops Terminal-Bench 4.0, so the picture is not uniformly bleak. For the model itself, see our Kimi K3.1 Pro open-weights coverage and the DeepSeek V4.1 Flash 1017 release.
One caution: this is a result of this particular blend. Shift weight toward coding-in-the-terminal and the open-weight picture improves, which is exactly the point critics make about composite scores.
The benchmarker's dilemma #
Nathan Lambert published a same-day reaction on Interconnects titled "Artificial Analysis v5 and the benchmarker's dilemma." The full piece sits behind a paywall, so we only cite its public opening, where he separates improving scores from designing good benchmarks: in his words, "Hill-climbing is easy, knowing what to hill-climb is hard." The article is framed as a review of the benchmarking industry and what is missing from it. We are not describing the paywalled arguments.
The dilemma shows up in v5 itself. Private test sets are the main defense against contamination, the same problem behind the GSM1k and LMArena findings. But a benchmark that is 60% secret cannot be replicated by anyone outside Artificial Analysis, so readers must trust the grading, the judge models and the choice of weights. v4.2 already drew timing and transparency questions on Hacker News, and moving from 40% to about 60% private raises the same point harder.
What this means for what you build or pay #
A stack of measuring rulers showing how Intelligence Index v5 weights nine evaluations differently
- Model picking: treat the top three as a tie for general use and choose on cost per task. Sol at $1.27 versus Opus at $10.04 is a very large difference for four points.
- Effort budget: test your own workload at medium and high effort before paying for max.
- Open weights: if you self-host, v5 says released models are meaningfully behind on hard science and agentic work, so verify on your own tasks rather than trusting a 2025-era "gap is closing" narrative. Compare with theMozilla claim of a roughly four-month gap .
- Version hygiene: v5 scores are not comparable to v4.x scores. Check the version label on any chart you cite.
- Run your own eval: a small private test set of your real tasks beats any public composite.
Honest limitations #
- All scores, costs and rankings are Artificial Analysis's own figures; we did not re-run them.
- Some third-party aggregators show different rankings (for example compiled self-reports with GPT-5.6 Sol ahead). Those are not the v5 run and should not be mixed with it.
- Kimi K3.1 Pro's weights were unreleased at publication, so its 46 is not usable open-weight capability yet.
- We could only read the public opening of the Interconnects piece.
Related reading #
Scores and weights reflect Artificial Analysis's Intelligence Index v5 as published October 27, 2026. The live index may change; check artificialanalysis.ai before deciding.