🔍 Read the full analysis: The New AI Model Leader Is Gemini 4 Carbon, By One Point—Here’s The Caveat on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Gemini 4 Carbon leads the Artificial Analysis Intelligence Index v5 with 59 points, one ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The overall score masks uneven results across nine tests, and Google is initially offering Carbon only to a limited group of cyber defenders, with no public release date announced.
Google’s Gemini 4 Carbon ranks first on the Artificial Analysis Intelligence Index v5, scoring 59 points—one ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The lead is specific to this nine-test benchmark: Carbon comes first on four tests and trails competitors on five, while Google says access is initially limited to a subset of cyber defenders.
Artificial Analysis launched version 5 of its index on October 27, then placed Carbon at the top the following day, according to the supplied report. Claude Opus 5.5, at its maximum setting, scored 58; GPT-6 Astra at maximum scored 57; GPT-6.1 Sol at maximum scored 55. Google’s previous model, Gemini 4 Argon, scored 50. The ranking combines performance across nine tests, with four broad categories—knowledge work, coding, science and general tasks—each accounting for 25% of the total.
Carbon leads on hard reasoning and knowledge questions in HLE-Diamond, science workflows in a terminal, the penalty-based AA-Omniscience test and business workflows on AutomationBench-AA. Its reported results include 69% on HLE-Diamond, compared with 68% for Claude Opus 5.5, and 65% on Terminal-Bench-Science, versus 63% for GPT-6 Astra. It also scores 78% on AutomationBench-AA, fractionally ahead of Gemini 4 Argon.
The model does not lead every category. Claude models outperform it on office-style knowledge work, Claude Opus 5.5 leads on real-world coding, GPT-6 Astra leads on physics research, Kimi K3.1 Pro tops terminal tasks, and Meta’s Muse Wonder Full leads on document reasoning. Carbon’s reported cost is $7.76 per index task at standard prices. That is below Claude Opus 5.5’s $10.04, but above GPT-6 Astra’s $6.24 and GPT-6.1 Sol’s $1.27.
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 Win
The result offers a useful comparison, but it does not establish that Carbon is the best choice for every user or workload. Its one-point margin depends on the index’s particular tests and weighting. The different category leaders show why a single composite score may not predict which model will perform best on a company’s coding, document or office tasks.
Cost and access also limit what the ranking means in practice. The report lists an introductory API price of $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20, respectively. Those token prices are distinct from the reported $7.76 index-task cost. Organizations deciding whether to use Carbon will need to compare the cost of completing their own tasks, not just the leaderboard position. For ordinary users, there is no immediate product change while access remains restricted.
Popular right now
De'Longhi 15-Bar Pump Espresso Machine with Milk Frother & Temperature Control
Ninja CREAMi Scoop & Swirl 13-in-1 Ice Cream and Soft Serve Maker, NC701
Ninja DualZone Air Fryer 10 Qt Smart 6-in-1 Grey (DZ950)
As an affiliate, we earn on qualifying purchases.
How the New Index Scores Models
Artificial Analysis describes its index as a combined score from nine tests, rather than a general verdict on intelligence. Version 5 divides the tests among knowledge work, coding, science and general performance, weighting each category equally. The supplied report says 60% of the index uses private test sets or held-out answers, intended to reduce the chance that models have memorized test questions.
Many tests involve extended work: models may act in a terminal, prepare reports or modify code. This makes the evaluation more focused on agent-like tasks than a short question-and-answer exchange, while also making runs more expensive. Results depend on the model and settings tested; the reported comparison uses specified configurations, including maximum settings for several competitors. A different benchmark mix could produce a different order.
Carbon’s AA-Omniscience result illustrates how test design affects rankings. It scores 54 on the test, six points ahead of Claude Opus 5.5, even though the source says several Claude and GPT models answer a higher share of questions correctly. Artificial Analysis’s measure penalizes wrong answers: Carbon’s reported hallucination rate is 11%, compared with about 50% for Claude Opus 5.5 and GPT-6 Astra. The index rewards restraint on this test, not simply the number of correct answers.
“choose by task and by cost per finished task, not by rank.”
— Thorsten Meyer, source report
Release Timing and Real-World Fit
No public launch date is provided for Gemini 4 Carbon. The stated rollout order is cyber defenders, trusted partners, then developers, businesses and consumers, but the timing and eligibility rules for each stage remain unspecified.
The supplied material reports index scores and prices but does not provide the full benchmark methodology, confidence intervals or independent replications of each result. It is also unclear how the index’s task-cost figure translates to particular customer workloads. The results should be read as a comparison on this benchmark, not as a guarantee of performance in other settings.
The source text ends partway through a discussion of the Coding Agent Index, so it does not supply a complete account of that comparison or Google’s full response. The available figures say Claude Code with Claude Opus 5.5 scored 62, while Google’s Antigravity CLI with Carbon scored 59; those are tool-and-model results, separate from the model-only index.
The Rollout and Further Testing
Google’s stated plan is to give Carbon first to a subset of cyber defenders, followed by trusted partners and then wider groups including developers, businesses and consumers. Google has not announced dates for those stages. Until access expands, companies and independent evaluators cannot broadly test the model against their own work.
Readers and prospective customers can watch for public availability, fuller pricing details and additional comparisons across tasks. For organizations, the practical next step will be to test Carbon on representative work and compare both accuracy and the cost of completed tasks with existing models. The leaderboard establishes a narrow lead on Artificial Analysis v5; broader performance and availability remain to be seen.
Key Questions
What does Gemini 4 Carbon’s first-place score mean?
It scored 59 points on Artificial Analysis’s Intelligence Index v5, one point ahead of Claude Opus 5.5. The score reflects nine benchmark tests and their weighting, not every possible use of an AI model.
Is Gemini 4 Carbon available to the public?
No public release is described in the source. Google is initially providing it to a subset of cyber defenders, with later access planned for trusted partners and broader groups. No dates are given.
Does Carbon beat other models on every test?
No. Carbon leads on four of the nine tests and loses on five. The source reports that Claude models lead office-style knowledge work and real-world coding, while other models lead some science, terminal and document tasks.
Why does Carbon win the AA-Omniscience test?
The test penalizes wrong answers, and Carbon reportedly answers cautiously. Its reported hallucination rate is 11%, though several competing models have higher plain-answer accuracy on the questions.
How much does Carbon cost?
The source lists an introductory API price of $2 per million input tokens and $10 per million output tokens, later increasing to $4 and $20. It separately reports a standard-price cost of $7.76 per index task; actual costs depend on usage.
Source: ThorstenMeyerAI.com
