What the Artificial Analysis Intelligence Index Measures, and Why Scores Move

The Artificial Analysis Intelligence Index is a weighted average of ten evaluations, grouped into four categories, that the company Artificial Analysis publishes to rank AI models. Headlines call it an "AI intelligence score". It is not a single exam and it is not an IQ. As of 14 October 2026 its methodology page describes version 4.3.2, Claude Opus 5.5 leads at about 58 out of 100, and the index has been rebuilt five times since 6 August, so a score quoted without its version is close to useless. Our earlier piece on what one "most intelligent" claim checked out to covered a specific announcement. This one opens the number itself: what it is built from, how it has changed, who leads it today and what gets lost when ten evaluations are folded into one score.

Ten evaluations, four categories, one weighted average

Artificial Analysis says version 4.3.2 incorporates ten evaluations and calculates the index as a weighted average across four categories: agents at 30% of the total, general knowledge at 30%, coding at 20% and scientific reasoning at 20%. Inside those categories each evaluation carries its own weight, and two of them, AA-Briefcase and AA-Omniscience, carry 15% each, half as much again as Humanity's Last Exam or CritPt at 10%. The suite is, in the publisher's words, primarily text-based and English-language.

• Agents, 30%: AA-Briefcase v1.1 (15%, 91 tasks), GDPval-AA v2.1 (10%, 220 tasks), AutomationBench-AA (5%, 657 tasks)
• Coding, 20%: Terminal-Bench 4.0 (10%, 66 tasks), SciCode (10%, 288 subproblems)
• General, 30%: AA-Omniscience (15%, 6,000 questions, split between accuracy and not hallucinating), GDP.pdf (10%, 100 tasks), AA-LCR v1.1 (5%, 100 questions)
• Scientific reasoning, 20%: Humanity's Last Exam (10%, 2,158 text-only questions), CritPt (10%, 70 tasks)

That weighting is a choice, not a law of nature. An index built to reward finishing multi-step agentic tasks will rank models differently than one built to reward graduate-level scientific recall, and the current version gives agents half as much weight again as scientific reasoning (30% against 20%). Since v4.3, 45% of the weight sits on evaluations whose questions or answers are private, so outsiders cannot check them. The methodology has been revised repeatedly, so a model's rank today is not comparable with a rank read off an older version of the same-named index.

How has the Intelligence Index changed in 2026?

Artificial Analysis lists every version on its methodology page and has published announcement articles for most of them. The changes are not cosmetic: evaluations were added and dropped and the weights were rebalanced.

• v4.0, January 2026: added GDPval-AA, AA-Omniscience and CritPt; removed MMLU-Pro, LiveCodeBench and AIME 2025; four categories at 25% each
• v4.1, 15 June 2026: GDPval-AA v2 introduced, Terminal-Bench Hard replaced by Terminal-Bench 2.1, tau2-Bench Telecom replaced by tau3-Banking, IFBench removed; weights agents 34%, coding 24%, scientific reasoning 24%, general 18%
• v4.1.1, 6 August 2026: tau3-Banking moved to a newer upstream release and the grader for several evaluations upgraded
• v4.2, 4 September 2026: AA-Briefcase (15%) and GDP.pdf (10%) added, GPQA Diamond removed, weights rebalanced; the article said 40% of the weight was now on private, held-out test sets
• v4.3, 7 September 2026: tau3-Banking replaced by AutomationBench-AA (5%) and Terminal-Bench 2.1 by Terminal-Bench 4.0; the private-set share rose to 45%
• v4.3.1 and v4.3.2, September 2026: patch releases with no announcement article, refreshing the judge panels and re-anchoring GDPval-AA, now v2.1

The effect on scores is large. Claude Fable 5.1 (maximum effort) scored 66 in Artificial Analysis's 1 September article. Under v4.3 it scores 53. Claude Opus 5 went from 63 to about 51. For the v4.1.1 change the publisher had said most models move by less than a point; the later versions moved leading models by more than ten. Any page that quotes a score of 60 or more for a 2026 model is quoting an older version, and the company says v5 is in progress but it had not been released when we read the site on 14 October.

Who leads the Artificial Analysis Intelligence Index now?

We read the leaderboard on 14 October 2026. Each model is listed at several reasoning-effort settings, and each setting is its own score, so the table below gives the best setting for each model. The company tracks more than 250 models.

• Claude Opus 5.5 (max, with fallback), Anthropic: 57.6, shown as 58
• Claude Sonnet 5.5 (max, with fallback), Anthropic: 56.0
• Muse Wonder Full (max), Meta: 55.8, available only inside Meta's Muse Code and not priced per token
• Claude Fable 5.1 (max, with fallback), Anthropic: 53.35
• GPT-6 Astra (max), OpenAI: 52.67
• Gemini 4 Argon (high), Google: 52.56, not publicly available when Artificial Analysis published its results on 30 September

Below them come Muse Wonder (52.4), GPT-6.1 Sol (51.8), Muse Spark 1.3 (48.1) and Grok 4.7 (46.5). The best open-weights model is MiMo-V2.6-Pro from Xiaomi at 46.3, ahead of GLM-5.3 (44.8), Kimi K3 (43.6), GLM-5.3-Flash (41.8) and Qwen3.8 (39.9). Two caveats come with the table. Anthropic rows are labelled "with fallback", meaning safety-flagged requests are routed to other models; Artificial Analysis reported that for Fable 5.1 this served about 4% of output tokens across the index. And the company publishes no interval for the composite score, so we cannot say how much of a one-point gap, such as the one between Fable 5.1, Astra and Argon, is real. We found confidence intervals only for the Elo-rated evaluations. For a different ruler, Epoch AI's Capabilities Index puts Opus 5.5, GPT-6 Astra, GPT-6.1 Sol, Sonnet 5.5 and Fable 5.1 within overlapping 95% intervals, a statistical tie at the top. How the Epoch scale is built, and why its anchors are two models and not people, is covered in the Epoch Capabilities Index explained.

Is there an Artificial Analysis Agentic Index or Coding Index?

Not as stand-alone pages that we could find. The agents category is 30% of the Intelligence Index, and the company publishes a separate Coding Agent Index (version 1.5), an equal-weighted composite of three benchmarks that scores agent-and-model pairs rather than bare models. On 14 October Claude Code with Sonnet 5.5 led it at 68.4, ahead of Muse Code with Muse Wonder Full at 68.0 and Claude Code with Opus 5.5 at 66.0.

Why "agentic" tasks got the biggest single weight

GDPval-AA, one of the three agentic evaluations (version 2.1 in the current index), illustrates what that category actually tests. The underlying GDPval benchmark, released by OpenAI, is built from tasks drawn from the real work of industry professionals — an average of 14 years of experience each — across 44 occupations spanning the top nine sectors that contribute most to U.S. GDP: finance, healthcare, law, software, and others. Its 220-task gold subset asks a model to produce an actual work product — a slide deck, a legal memo, a spreadsheet model, a diagram — graded by having human experts compare the model’s output against a professional’s, rather than checking a single right answer. OpenAI’s own reporting on the benchmark states that frontier models complete these tasks roughly 100 times faster and cheaper than the human professionals whose work anchors the comparison, though speed and cost are not the same thing as quality of output.

That is a meaningfully different thing to measure than a written exam. A model can be excellent at recalling and applying specialist knowledge under exam conditions and comparatively weak at the sustained, multi-step judgment a real deliverable requires, or the reverse. Weighting agentic, real-work evaluations at 30% — the joint-largest share, tied with general knowledge — is Artificial Analysis’s stated bet that economically useful task completion is the more important thing to track. That is a defensible editorial choice. It is also a choice, and a different one would move the leaderboard without any model actually changing.

The hardest ingredient: an exam built to resist being solved

Humanity’s Last Exam is one of the two scientific-reasoning evaluations, worth 10% of the index, and it exists because the benchmarks before it stopped working. The index scores a text-only subset of 2,158 questions from the May 2025 revision of the 2,500. The Center for AI Safety and Scale AI released it publicly in January 2025, built from roughly 2,500 questions spanning more than 100 subjects — about 41% mathematics, with roughly 14% of the set incorporating multi-modal elements such as images or diagrams, and the remainder split across physics, biology, medicine, computer science and the humanities — contributed by close to 1,000 subject-matter experts across more than 500 institutions in 50 countries. Submitted questions were tested against leading AI models first; only ones the models answered incorrectly, or did no better than random guessing on, went through two further rounds of human expert review before making the final set, with a $500,000 prize pool distributed to the strongest submissions. Stanford HAI’s AI Index 2025 report cites the project as a direct response to benchmark saturation — frontier models were already scoring close to the ceiling on the popular benchmarks that came before it. How that exam was built, and how its answer key has been audited, is covered in our guide to the Humanity's Last Exam leaderboards.

• GPT-4o (2024): 2.7%
• Claude 3.5 Sonnet (2024): 4.1%
• OpenAI o1: 8%
• Gemini 3.1 Pro and Claude Opus 4.6 (later rounds): roughly 40–50%
• Artificial Analysis's text-only version, 14 October 2026: Claude Opus 5.5 at 61.4%, Claude Fable 5.1 at 59.1% and GPT-6 Astra at 54.7%

What one composite number hides

A single index score invites a mistake this site has already covered in a different context. Our explainer on full-scale IQ versus index scores makes the point that a single full-scale number is really an average across several separate measurements, and that two of those measurements can pull apart on the same person’s report even when the top-line score looks unremarkable. The same structure sits inside an AI intelligence index: two models can land on an identical overall score while one is well ahead on agentic task completion and meaningfully behind on scientific reasoning, or the reverse. The composite does not show which — and unlike a WISC-V or WAIS report, there is no equivalent of the confidence-interval band this site’s piece on IQ score margin of error covers, printed next to the headline AI figure for a general reader to see.

It is also worth being precise about what these scores are not. Our piece on why an AI’s score on an IQ-style question set is not an IQ covers the deeper reason: an IQ score is anchored to a norm — the distribution of scores in an actual human reference population, expressed as a percentile against real people who sat the same test under the same conditions. An intelligence index score has no such population behind it. It is a percentage or a points total on a fixed set of questions, comparable only to another model’s score on that identical set, not to a percentile ranking against anyone.

Why the same two models get reported differently by different sources

None of this means the underlying numbers are arbitrary, but it does mean that the same model gets different numbers from different rulers. On Humanity's Last Exam, Artificial Analysis shows Fable 5.1 at 59.1% and GPT-6 Astra at 54.7%, while Scale AI's own text-only leaderboard, which uses a different automatic judge and its own question set, shows Astra first at 54.2% and Fable 5.1 at 46.8%. The published methodology of each is sound on its own terms; they are not the same measurement. We found no confidence interval for the composite index on any Artificial Analysis page: its only interval statements concern the Elo ratings of two evaluations, GDPval-AA and AA-Briefcase. Our piece on why the same model gets different benchmark scores from different labs covers a related cause of disagreement, grading-rule choices inside a single benchmark, which alone can move one model's reported score by dozens of points.

Next time a leaderboard number gets reported as "the most intelligent AI," the useful question is not which company published it. It is which ten evaluations are sitting underneath the average, and how much weight each one was given before anyone saw the final digit. Our own IQ test reports a percentile against a real population for exactly this reason — a single number is only as meaningful as what stands behind it. METR publishes a different AI measure, a task-length time horizon with wide error bars, explained in METR time horizon explained.

Common questions