Artificial Analysis Intelligence Index v5: 60% private tests, three evaluations dropped, and Gemini 4 Carbon on top at 59

Artificial Analysis Intelligence Index v5 is the new version of the composite score that independent evaluator Artificial Analysis uses to rank frontier models, launched on October 27, 2026. The firm calls it "our biggest upgrade since v4 launched in January." It swaps out three evaluations, adds two, and says "60% of the Index is now based on evaluations with private test sets or held-out answers." At launch, Claude Opus 5.5 led at 58. A day later, Artificial Analysis scored Google's Gemini 4 Carbon at 59, which Google quoted in its Carbon announcement. Every score here is Artificial Analysis's, not 9to5AI's, and none can be compared with the v4.3.2 numbers in our earlier stories.

This report is based on Artificial Analysis's October 27 v5 launch article, its methodology page, its September 7 v4.3 article, its October 28 Gemini 4 Carbon article and model page, and its public leaderboard, all checked October 29, 2026.

The short version

• Nine evaluations, four categories at 25% each: knowledge work, coding, science and general knowledge.
• Added: FrontierCode 2.0 and Terminal-Bench-Science. Removed: GDPval-AA, SciCode and AA-LCR. Upgraded: HLE-Diamond, AA-Omniscience v1.1 and CritPt v2.
• Private share: 60% of the weight, up from 45% in v4.3.
• Who leads: Gemini 4 Carbon (high) at 59, per Artificial Analysis's October 28 article, then Claude Opus 5.5 (max) at 58, GPT-6 Astra (max) at 57 and GPT-6.1 Sol (max) at 55. Carbon isn't publicly available, but the public leaderboard lists it first as of October 29.
• Not comparable with v4.3.2: several models moved by 3 to 9 points with the methodology change, not because the models changed.

What is the Artificial Analysis Intelligence Index?

It's a single score that Artificial Analysis computes by running a set of evaluations on each model and taking a weighted average. The firm chooses the evaluations and the weights, runs them itself, and publishes the results with the effort setting used for each model, such as "max" or "high." Model makers cite it: Google put Carbon's v5.0 score at the top of its launch table.

Artificial Analysis describes the index as "a primarily text-based, English-language evaluation suite" and says that, "like all evaluation metrics, it has limitations and may not apply directly to every use case."

What changed in Intelligence Index v5?

The categories and weights changed, and so did a third of the evaluations.

Last verified: October 29, 2026

The categories. In v4.3, Artificial Analysis weighted Agents at 30%, Coding 20%, General 30% and Scientific Reasoning 20%. In v5.0, four categories each get 25%:

• Knowledge Work: AA-Briefcase 15%, AutomationBench-AA 5%, GDP.pdf 5%.
• Coding: FrontierCode 2.0 15%, Terminal-Bench 4.0 10%.
• Science: Terminal-Bench-Science 15%, CritPt v2 10%.
• General: AA-Omniscience v1.1 15% (10% accuracy, 5% non-hallucination), HLE-Diamond 10%.

The additions. FrontierCode 2.0 is Cognition's private benchmark of 132 real open-source repository tasks, graded on whether a maintainer would merge the change; Artificial Analysis says "the tasks stay private to avoid contamination." Terminal-Bench-Science has 70 agentic scientific research tasks done in a terminal.

The upgrades. HLE-Diamond is a 1,000-question version of Humanity's Last Exam, split evenly between reasoning and knowledge questions. AA-Omniscience v1.1 drops saturated and publicly released questions and adds new ones; scores average 2 points lower than in v1. CritPt v2 is a private 40-question hard subset, run with a code-execution tool. Artificial Analysis says the 30 questions it removed were ones seven frontier models solved in all five attempts at launch.

The removals. The v5 article lists GDPval-AA v2.1, SciCode and AA-LCR v1.1 as removed but doesn't say why. The methodology page says GDPval-AA, which was in the index "up to and including v4.3.2," is now a standalone evaluation.

Also new: a Progress Index v1, which Artificial Analysis says "places language models released since November 2022 on one capability scale." The firm says the Intelligence Index "remains our primary measure of current model intelligence."

Who leads the Intelligence Index v5?

It depends on whether you count a model the public can't use yet.

Artificial Analysis Intelligence Index v5.0, top models

Index score (higher is better), configuration as labeled by Artificial Analysis

Gemini 4 Carbon (high), not public59

Claude Opus 5.5 (max, with fallback)58

GPT-6 Astra (max)57

GPT-6.1 Sol (max)55

Claude Sonnet 5.5 (max, with fallback)54

GPT-6.1 Terra (max)51

Gemini 4 Argon (high)50

Show the data

Model	Score
Gemini 4 Carbon (high), not public	59
Claude Opus 5.5 (max, with fallback)	58
GPT-6 Astra (max)	57
GPT-6.1 Sol (max)	55
Claude Sonnet 5.5 (max, with fallback)	54
GPT-6.1 Terra (max)	51
Gemini 4 Argon (high)	50

At launch, October 27. Artificial Analysis wrote: "Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra (max, 57) and GPT-6.1 Sol (max, 55)," and "Anthropic and OpenAI hold the top five models." Gemini 4 Argon (high) scored 50. On the public leaderboard, checked October 29, Claude Sonnet 5.5 (max) scores 54 and GPT-6.1 Terra (max) 51. Opus 5.5 was released on September 22, 2026, GPT-6 Astra on September 3 and GPT-6.1 Sol on September 29, according to their makers.

From October 28. Artificial Analysis published "Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5," scoring Carbon (high) at 59, with the top score on four of the nine evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. Google announced Carbon the same day. But Artificial Analysis also says Carbon "is being rolled out to selected users and is not yet publicly available," and doesn't say how it got access. By October 29, its public leaderboard and its index page listed Carbon (high) first at 59, ahead of Opus 5.5 at 58. Our Gemini 4 Carbon vs Opus 5.5 vs GPT-6.1 Sol comparison covers what you can actually buy.

Open-weights models fell further behind. Artificial Analysis says the best, Kimi K3.1 Pro, "scores 46, 12 points behind the leader," against a 3-point gap in v4.3; its weights hadn't been released yet. Among models with released weights, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) top out at 32, below Anthropic's small Claude Haiku 5.5 (max) at 36.

Which model leads each evaluation?

Artificial Analysis says "no single model leads them all." Its October 27 article named these leaders:

• FrontierCode 2.0: Opus 5.5 at 46%; GPT-6 Astra, GPT-6.1 Terra and Sonnet 5.5 at 42%.
• Terminal-Bench-Science: GPT-6 Astra 63%, Opus 5.5 59%, GPT-6.1 Sol 58%. Most open-weights models score under 10%.
• HLE-Diamond: Opus 5.5 68% (78% on reasoning questions, 58% on knowledge questions), GPT-6 Astra 64%.
• CritPt v2: GPT-6 Astra (high) 72%, GPT-6.1 Sol (max) 69%, Opus 5.5 61%.
• AA-Omniscience v1.1: Opus 5.5 at 48.
• AA-Briefcase: Sonnet 5.5 at 1812 Elo.
• Terminal-Bench 4.0: Kimi K3.1 Pro at 67%.
• AutomationBench-AA: Gemini 4 Argon at 78%.
• GDP.pdf: Muse Wonder Full at 33% all-pass.

On October 28, Artificial Analysis added that Carbon scores 54 on AA-Omniscience, ahead of Opus 5.5's 48, and 64% on Terminal-Bench 4.0, second. Its hallucination rate rose to 11% from Argon's 4%, which the firm says is still the second-lowest of any model.

How much does it cost to run, and why does effort matter more now?

v5 is more expensive to run. Artificial Analysis says every model used more output tokens than in v4.3, with the median rising from 35,000 to 59,000 per task. The four agentic evaluations, Terminal-Bench-Science, Terminal-Bench 4.0, AA-Briefcase and FrontierCode 2.0, are 55% of the weight but a median of about 90% of the output tokens.

Its cost per index task, as Artificial Analysis calculates it from list prices: Opus 5.5 (max) $10.04, GPT-6 Astra (max) $6.24, GPT-6.1 Sol (max) $1.27, GPT-6.1 Terra (max) $0.96 and Claude Fable 5.1 (max) $13.04. Carbon costs $7.76, priced at Google's later $4/$20 rate rather than its $2/$10 launch price. Artificial Analysis says GPT-6.1 Sol "scores within 4 points of Claude Opus 5.5 (max) at about an eighth of the cost per task."

Effort settings also spread scores further. Opus 5.5 scores 35 at low effort and 58 at max. Artificial Analysis says Sonnet 5.5's gap between low and max grew from 20 points to 29, and GPT-6.1 Sol's from 10 to 20. A headline score is only as meaningful as the effort label next to it.

Can I compare v5 scores with older Intelligence Index numbers?

No. Earlier 9to5AI stories quoted version 4.3.2 or 4.3, which used different evaluations and weights. The version history on Artificial Analysis's methodology page dates v5.0 from "October 2026 to current." It describes v4.3.1 and v4.3.2, used in September and October 2026, as changes to judge panels and rating anchors, not to the list of evaluations or their weights. Changes between v4.3.x and v5.0 reflect the new method, not changes in the models:

Last verified: October 29, 2026

Those older scores appear in our Argon vs Opus 5.5 vs GPT-6 comparison, our GPT-6 Astra vs 6.1 Sol vs Luna explainer, our Kimi K3.1 Pro story and our Muse Wonder story. They were correct for their version on the dates shown and should be read as v4.3.x results.

What are the limits of the Intelligence Index v5?

• One firm's choices. Artificial Analysis picks the nine evaluations and sets the weights. A different mix would give a different order, and the four best-scoring public models, Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and Sonnet 5.5, sit within four points.
• Private tests cut both ways. Private sets reduce the chance that models were trained on the answers, which is the firm's stated aim, but outsiders can't inspect 60% of the index.
• Small subsets. CritPt v2 is 40 questions, so a few answers move a model's score.
• Configurations differ. Scores are tied to effort labels such as "max," "high" and "with fallback," which Artificial Analysis uses for Claude runs, and the top models weren't all run at the same setting.
• Text and English only, by the firm's own description.
• Moving parts. AA-Briefcase scores are an Elo rating that Artificial Analysis says is "frozen at the time of a model's addition," and the firm says some reference parameters may be updated over time.
• Unreleased models. Carbon's 59 is for a model the public can't use, obtained in a way Artificial Analysis doesn't describe.

What's confirmed, reported and uncertain

• Confirmed (Artificial Analysis's own pages): the v5.0 evaluations, weights and 60% private share; the added, upgraded and removed evaluations; the October 27 scores; Carbon's 59 on October 28 and its first place on the public leaderboard by October 29; the version history; the cost and token figures.
• Reported by vendors: release dates for the models, from Anthropic, OpenAI and Google.
• Uncertain: why GDPval-AA, SciCode and AA-LCR were dropped; how Artificial Analysis accessed Carbon; when Carbon will be publicly available.