🔍 Read the full analysis: Gemini 4 Carbon Edges Ahead By One Point—Here’s What The AI Ranking Misses on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
Gemini 4 Carbon scored 59 on Artificial Analysis’ Intelligence Index v5, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. It leads four of nine tests, but trails competitors on other tasks, and Google has not yet made it publicly available.
Google’s Gemini 4 Carbon took the top spot on Artificial Analysis’ Intelligence Index v5 with 59 points, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The result puts Carbon first on this nine-test benchmark, but does not establish it as the best model for every task; Google says access is currently limited to a subset of cyber defenders.
Artificial Analysis launched the rebuilt index on October 27. Its scores place Gemini 4 Carbon, using a high-reasoning setting, at 59; Claude Opus 5.5 at 58; GPT-6 Astra at 57; and GPT-6.1 Sol at 55. Google’s previous model, Gemini 4 Argon, scored 50. Those are results on this index and these model settings, not a universal measure of AI capability.
Carbon ranked first on four of the nine tests: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. It did not lead the other five, which cover areas including office-style knowledge work, coding, physics research, terminal tasks and document reasoning. For example, Claude Opus 5.5 scored 46% on FrontierCode 2.0, compared with Carbon’s 41%, while Claude Sonnet 5.5 and Opus 5.5 topped the reported AA-Briefcase results.
The index’s cost estimate is $7.76 per task at standard prices, according to the source report. That is below Claude Opus 5.5’s listed $10.04 but above GPT-6 Astra’s $6.24 and GPT-6.1 Sol’s $1.27. The report also lists introductory API pricing for Carbon at $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20, respectively. The exact circumstances and timing for those prices were not independently established in the supplied material.
Infographic 01 · What the #1 is made of
Number one overall. First in four of nine tests.
The Intelligence Index is a decathlon score: nine tests, one number. Carbon wins on reasoning, science and careful knowledge, and trails on office work and coding.
The nine tests behind the score
Carbon at its highest setting (high) against the best model on each test.
| Test | Weight | Carbon | Leader or runner-up |
|---|---|---|---|
| HLE-Diamond Hard reasoning + knowledge questions | Weight 10% | Carbon 69% | Carbon leads Next: Claude Opus 5.5, 68% |
| Terminal-Bench-Science Science research work in a terminal | Weight 15% | Carbon 65% | Carbon leads Next: GPT-6 Astra, 63% |
| AA-Omniscience v1.1 Knowledge, wrong answers penalized | Weight 15% | Carbon 54 | Carbon leads Next: Claude Opus 5.5, 48 |
| AutomationBench-AA Business workflows across apps | Weight 5% | Carbon 78% | Carbon leads Next: Gemini 4 Argon, 78% |
| AA-Briefcase Office-style knowledge work | Weight 15% | Carbon 1,609 Elo | Behind Claude Sonnet 5.5, 1,812 |
| FrontierCode 2.0 Code a maintainer would merge | Weight 15% | Carbon 41% | Behind Claude Opus 5.5, 46% |
| CritPt v2 (hard) Research-level physics | Weight 10% | Carbon 63% | Behind GPT-6 Astra, 71% |
| Terminal-Bench 4.0 Agentic terminal tasks | Weight 10% | Carbon 65% | Behind Kimi K3.1 Pro, 67% |
| GDP.pdf Reasoning over documents | Weight 5% | Carbon 28% | Behind Muse Wonder Full, 33% |
The surprise in the knowledge test
- Accuracy: Claude Fable 5.1 68%, Claude Opus 5.5 66%, GPT-6.1 Sol and GPT-6 Astra 62%, Carbon 58%
- Hallucination rate: Carbon 11%, Claude Opus 5.5 and GPT-6 Astra about 50%
- Carbon wins because it rarely bluffs, not because it knows the most
Read the weights
- Each of the four blocks counts 25%: knowledge work, coding, science, general
- Carbon trails in two of the four heaviest tests (15% each): office-style work and real-world code
- Hallucination rate = wrong answers among all questions not answered correctly
Source: Artificial Analysis, 28 Oct 2026; Index v5 announcement, 27 Oct 2026; Artificial Analysis model data, retrieved 29 Oct 2026. Scores rounded as in the article; Carbon tested at high, other models at their highest tested setting.
Powered by Thorsten Meyer AI
A Narrow Lead, Not a Universal Winner
The ranking may influence which models companies test or buy, but its one-point margin is small, and task-level results vary. Organizations using AI for software development, office work or research may get a different answer from the overall score than teams whose work resembles Carbon’s strongest tests. The benchmark is a starting point for evaluation, not a substitute for trials on a company’s own tasks.
Cost also changes the practical comparison. The source report’s per-task estimates put Carbon below Claude Opus 5.5 but above GPT-6 Astra and far above GPT-6.1 Sol. These figures concern the benchmark’s task mix; they do not establish what any particular business will pay for its own workload. Users should compare cost per completed task, quality and speed under their actual conditions.
For everyday users, the immediate effect is limited because Carbon is not publicly available, according to the report. If the staged rollout proceeds as described, its relevance will depend on access, product integration and performance on real use cases—not just its leaderboard position.
Popular right now
Samsung 42-Inch S90H OLED 4K Smart TV Bundle
Vitamix Propel Series 510 Professional-Grade Blender, 3 Programs, Black
Dyson V8 Cyclone Cordless Vacuum
As an affiliate, we earn on qualifying purchases.
How Index v5 Produces Its Score
Artificial Analysis describes the Intelligence Index as a combined score from nine tests, grouped into four equally weighted areas: knowledge work, coding, science and general performance. The supplied report says knowledge work includes AA-Briefcase, AutomationBench-AA and GDP.pdf; coding includes FrontierCode 2.0 and Terminal-Bench 4.0; science includes Terminal-Bench-Science and CritPt v2; and the general category includes AA-Omniscience and HLE-Diamond.
Many tasks involve long agent workflows, such as working in a terminal, producing reports or changing code. The report says 60% of v5 uses private test sets or held-out answers, intended to reduce the chance that models have memorized questions. The index therefore measures performance on a particular set of tasks and settings; different tests or weightings could produce a different ordering.
One result illustrates how test design affects rankings. Carbon scored 58% accuracy on AA-Omniscience, below several listed Claude and GPT models. Yet it led that test’s score because the metric penalizes wrong answers and Carbon’s reported hallucination rate was 11%, compared with about 50% for Claude Opus 5.5 and GPT-6 Astra. In this measure, fewer incorrect guesses can outweigh a higher number of correct answers.
“Choose by task and by cost per finished task, not by rank.”
— Thorsten Meyer, source report
Access and Real-World Results Remain Open
The supplied material does not include the full Artificial Analysis methodology, independent verification of the reported scores, or the underlying test results. It also does not specify how many models or runs were included, whether the reported differences are statistically meaningful, or how repeatable the one-point gap is. The rankings should be read as reported benchmark results, not proof of a broad capability advantage.
Google’s rollout is described as starting with a subset of cyber defenders, followed by trusted partners and then wider groups, but the report gives no release date. It is also unclear when API pricing will change from the introductory rates, what access conditions will apply, and whether public-facing Google products will use Carbon.
Some task-level comparisons in the source are incomplete, including a sentence about a separate coding-agent comparison that ends mid-quote. This article therefore does not infer the rest of that claim. The index also cannot establish how Carbon performs on every kind of work, or whether its benchmark cost estimate matches an individual user’s usage.
Watch for Access and Independent Testing
The next clear milestone is wider availability. The source report says Google plans to move from a limited group of cyber defenders to trusted partners, then developers, businesses and consumers, but has not supplied dates for those stages. Until access expands, most users cannot test Carbon directly against the alternatives.
Companies considering the model can compare it with rivals on representative internal tasks, recording accuracy, completion rate, latency and total cost. That is particularly relevant because the index shows different leaders for different categories, while the overall score compresses those differences into a single number.
Further updates may clarify Carbon’s API availability and pricing, and whether its measured performance holds up in independent or customer evaluations. For now, the confirmed development is a narrow first-place ranking on one benchmark—not a general finding that Carbon is the best choice for all users.
Key Questions
What is Gemini 4 Carbon’s score?
It scored 59 points on Artificial Analysis’ Intelligence Index v5, one point above Claude Opus 5.5 and two above GPT-6 Astra, according to the supplied report.
Does the ranking mean Carbon is the best AI model?
No. It ranks highest on this nine-test index. Carbon led four tests and did not lead five, so the best fit depends on the task and how the model performs under real conditions.
Can the public use Gemini 4 Carbon now?
Not according to the source report. Google is said to be giving access first to a subset of cyber defenders, with later stages planned but no dates provided.
Why did Carbon lead the knowledge test despite lower accuracy?
On AA-Omniscience, the score penalizes incorrect answers. Carbon had lower reported accuracy than several models but an 11% hallucination rate, so fewer wrong guesses helped it lead the test’s overall score.
How much does Carbon cost?
The report estimates $7.76 per index task at standard prices and lists introductory API pricing of $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20. Availability and timing for those rates have not been clarified in the supplied material.
Source: ThorstenMeyerAI.com
