🔍 Read the full analysis: What The Latest AI Ranking Says—and Doesn’t Say—About Gemini 4 Carbon 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’s Intelligence Index v5, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The result reflects performance across nine tests, not a universal ranking: Carbon leads four tests, trails on five, and is currently being offered only to a small group of cyber defenders.
Google’s 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 ranking is a narrow result across nine selected tests, not a finding that Carbon is best at every kind of AI task; Google says access is currently limited to a subset of cyber defenders.
Artificial Analysis’s v5 leaderboard gives Claude Opus 5.5 (max) 58 points, GPT-6 Astra (max) 57, and GPT-6.1 Sol (max) 55. Google’s prior model, Gemini 4 Argon, scores 50. Carbon takes first place on four of the nine underlying tests, according to the source report, and does not lead the other five.
Its strongest results include 69% on HLE-Diamond, compared with 68% for Claude Opus 5.5, and 65% on Terminal-Bench-Science, compared with 63% for GPT-6 Astra. Carbon also scores 78% on AutomationBench-AA, a business-workflow test. It leads AA-Omniscience with 54 points, but that score rewards avoiding incorrect answers as well as answering correctly.
The cost figures in the report put Carbon at $7.76 per index task at standard prices. That is below Claude Opus 5.5 at $10.04, but above GPT-6 Astra at $6.24 and substantially above GPT-6.1 Sol at $1.27. The report also lists an introductory API price of $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20 respectively. Pricing and task-cost comparisons depend on the stated model settings and benchmark workload.
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 With Real Trade-Offs
The result matters because enterprise buyers often use leaderboards to shortlist models, while a one-point gap can look more decisive than it is. Carbon’s total score summarizes a particular mix of tasks and weights; it does not establish that the model will perform best on a company’s own documents, codebase, or workflow. The reported gaps on individual tests are also small in some cases, and the supplied material does not provide uncertainty intervals that would show how stable those differences are.
Costs and access complicate the practical picture. At the benchmark’s reported standard rates, Carbon costs less per index task than Claude Opus 5.5 but more than GPT-6 Astra and GPT-6.1 Sol. Google has not made it generally available, so most developers and businesses cannot yet verify the model against their own needs. For now, the score is a signal for evaluation, not a reason by itself to switch products.
Products Worth Considering
Logitech MX Brio Ultra HD 4K Webcam, Graphite
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black
Samsung Galaxy Tab S10 FE 128GB Wi-Fi Android Tablet, Blue
As an affiliate, we earn on qualifying purchases.
How Index v5 Builds Its Score
Artificial Analysis launched Index v5 on October 27, describing it as a rebuilt leaderboard that combines nine tests into one score. The tests are divided into four equally weighted areas: knowledge work, coding, science, and general performance. Individual test weights differ within each area; for example, AA-Briefcase and FrontierCode 2.0 each account for 15% of the overall score.
The source report says 60% of the index uses private test sets or held-out answers, intended to reduce the effect of models memorizing questions. Many tasks also require longer agent-style work, such as using a terminal, preparing reports, or fixing code. Those design choices aim to test more than short question-and-answer performance, but the resulting score still depends on the benchmark’s task selection, weighting, and model configuration.
On some measures, Carbon’s first-place result comes from restraint rather than the highest raw knowledge accuracy. The report gives Carbon 58% accuracy on AA-Omniscience, below several Claude and GPT models, but says its hallucination rate on that test is 11%, versus about 50% for Claude Opus 5.5 and GPT-6 Astra. Artificial Analysis defines that rate as the share of wrong answers among questions the model did not get right. The distinction matters: a scoring system that penalizes wrong answers can favor a model that abstains more often, even when another model answers more questions correctly.
What the Score Leaves Open
The supplied material does not include the full leaderboard methodology, test-by-test results for every competitor, or statistical uncertainty around the one-point overall gap. It is therefore unclear whether the difference between Carbon and Claude Opus 5.5 would persist across repeated runs or other evaluation settings. A ranking can also change if the benchmark’s tests or weights change.
Several limitations remain relevant to interpreting the result. Carbon loses on five of the nine tests, including office-style knowledge work and real-world coding, according to the report. It also takes longer to reason, although the supplied material provides no timing figures. The reported task costs are benchmark-specific and do not establish what a customer will pay for a different workload. Finally, Carbon’s public release date and the schedule for access beyond the initial cyber-defender group are not specified.
Access and Independent Testing
Google’s stated rollout order is cyber defenders first, then trusted partners, followed by developers, businesses, and consumers. No dates are supplied for those later stages. Until access expands, independent users and organizations will have limited opportunity to reproduce the benchmark results or compare Carbon on their own tasks.
For companies considering the model, the next useful evidence will be tests on their own workloads that track accuracy, abstentions, latency, and cost per completed task. Readers should also watch for updates to Artificial Analysis’s leaderboard and any fuller disclosure of methodology or repeated-run results. For everyday users, the current ranking does not change what models are available to use today.
Key Questions
What did Gemini 4 Carbon score?
It scored 59 points on Artificial Analysis’s Intelligence Index v5, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra.
Does the ranking prove Carbon is the best AI model?
No. It ranks performance across nine selected tests with set weights. Carbon leads four tests and loses five, so the result does not establish that it is best for every task or user.
Why did Carbon lead the knowledge test if other models answered more correctly?
The report says Carbon had lower raw accuracy on AA-Omniscience than several rivals, but a much lower reported hallucination rate. Because the test penalizes wrong answers, answering cautiously can improve its score.
Can the public use Gemini 4 Carbon now?
Not according to the source report. Google is initially providing access to a subset of cyber defenders, with broader access planned for later groups but no public dates specified.
How much does Carbon cost?
The report lists an introductory API price of $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20. It also estimates $7.76 per index task at standard prices; actual costs may differ by workload.
Source: ThorstenMeyerAI.com
