🔍 Read the full analysis: Gemini 4 Carbon’s #1 AI Ranking: What A One-Point Margin Can’t Explain on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
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 on four of nine tests, but the ranking does not make it the top performer for every task, and Google has not yet made it publicly available.
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 result puts Carbon at the top of this nine-test benchmark, but it is not a general verdict that the model outperforms rivals on every kind of work: it leads four tests and trails on five, and Google has not yet released it publicly.
Artificial Analysis introduced version 5 of its index on October 27, then listed Carbon at the top the following day, according to the source report. The reported scores are 59 for Gemini 4 Carbon, 58 for Claude Opus 5.5, 57 for GPT-6 Astra, 55 for GPT-6.1 Sol and 50 for Google’s previous model, Gemini 4 Argon. Carbon’s score is for its high-reasoning setting; the rival scores cited are for their maximum settings.
The index combines results from nine tests across knowledge work, coding, science and general capabilities, with each category carrying equal weight. Carbon ranks first on HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. It does not lead the office-work or real-world coding measures, among others. On HLE-Diamond, it scored 69% against Claude Opus 5.5’s 68%; on Terminal-Bench-Science, it scored 65% against GPT-6 Astra’s 63%.
The cost and release status complicate comparisons. The source gives Carbon’s standard price for an index task as $7.76, compared with $10.04 for Claude Opus 5.5 and $6.24 for GPT-6 Astra; the task-cost method and assumptions are not detailed in the material provided. Google’s stated introductory API prices are $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20. Those prices do not mean users can access the model now: Google is starting with a subset of cyber defenders, with no public release date supplied.
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 One-Point Lead Has Limits
For businesses choosing an AI model, the result offers a useful comparison, not a universal buying recommendation. The index’s equal weighting across four categories means a model’s combined position can conceal clear differences between tasks. Carbon’s reported strength in science workflows and cautious answers may appeal in some settings, while its lower result on office-style knowledge work and real-world coding may matter more to other teams.
The ranking also illustrates why a score needs its test conditions. Artificial Analysis says 60% of the index uses private test sets or held-out answers, reducing the chance that models have simply memorized test questions. Many tasks involve longer agent workflows, such as working in a terminal or producing reports. That makes the benchmark relevant to practical use, but the final score still represents one model configuration on nine selected tests, not every product feature, workflow or user experience.
Cost comparisons can change the decision as well. The source’s per-task estimates put Carbon below Claude Opus 5.5 but above GPT-6 Astra and substantially above GPT-6.1 Sol. These are reported benchmark-task costs, not a guaranteed bill for a company’s own usage. Organizations will need to test their workloads and measure the cost of completed work, including retries and human review, before treating the leaderboard as a procurement guide.
Products Worth Considering
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
Apple iPad Air (4th Generation, 10.9-inch, Wi-Fi, 256GB) – Space Gray
As an affiliate, we earn on qualifying purchases.
How the Index Scores Models
Artificial Analysis describes Intelligence Index v5 as a combined score from nine benchmarks grouped into four equally weighted blocks: knowledge work, coding, science and general tasks. The individual tests do not count equally within each block. For example, AA-Briefcase accounts for 15% of the full index, as do FrontierCode 2.0 and Terminal-Bench-Science. The overall score reflects performance across this selected mix rather than a single exam.
Carbon’s AA-Omniscience result highlights how a benchmark’s scoring rule can shape a ranking. The source reports that Carbon answered 58% of questions correctly, below several Claude and GPT models on raw accuracy. But the test 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 that test, a model that declines to guess can score better than one that answers more questions but gets more of them wrong. The report also says Gemini 4 Argon’s hallucination rate was 4%, so Carbon was not the most cautious model in that comparison.
Performance varies on other measures too. In AA-Briefcase, the source reports Carbon at 1,609 Elo, behind Claude Sonnet 5.5 at 1,812 and Claude Opus 5.5 at 1,800. In FrontierCode 2.0, Carbon scored 41% against Claude Opus 5.5’s 46%. These results help explain why first place overall does not mean first in each category.
What the Ranking Leaves Open
The source material does not include the full Artificial Analysis methodology, all nine test results, or the complete account of how the score is calculated. It also ends partway through a statement about Google’s coding-agent result, so that claim cannot be reported in full here. Readers should treat the stated scores and comparisons as figures attributed to the report, rather than independently verified measurements.
It is also unclear when Carbon will reach developers, businesses or consumers. Google’s reported rollout begins with selected cyber defenders and proceeds to trusted partners and broader groups, but the source gives no dates for those stages. The provided material does not specify how the reported task-cost estimates were calculated or how results might change under different prompts, tools or model settings.
Finally, a one-point gap does not by itself show whether the lead is statistically meaningful or stable across repeated runs. The source gives no uncertainty range or significance analysis for the overall scores. Nor does this index answer questions about reliability, privacy, latency or performance on a particular company’s own work.
Access and Testing Ahead
According to the source, Google plans to begin with a subset of cyber defenders, then extend access to trusted partners, developers, businesses and consumers. No schedule was supplied for those stages, so the immediate next milestone is wider availability rather than a dated public launch.
Until then, companies can use the benchmark as a shortlist signal, not a substitute for evaluation. Once access is available, they can compare Carbon with alternatives on representative tasks, track accuracy and completion rates, and calculate the cost of finished work. For everyday users, the ranking does not change which Google model is available today; the practical effect depends on when Carbon is released and how it performs in products people can actually use.
Key Questions
What does Gemini 4 Carbon’s number-one ranking mean?
It means Carbon scored 59 on Artificial Analysis Intelligence Index v5, one point above Claude Opus 5.5. It is first on that benchmark’s selected mix of nine tests, not necessarily the best model for every task.
Which tests did Carbon lead?
The source says Carbon ranked first on HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. It trailed on five other tests in the index.
Is Gemini 4 Carbon publicly available?
No public availability is reported. Google is starting with a subset of cyber defenders and plans later access for partners, developers, businesses and consumers, but no release dates were provided.
Does Carbon know more than the other models?
Not according to the raw knowledge-accuracy figures in the source. Several Claude and GPT models answered a larger share of questions correctly. Carbon led AA-Omniscience because the test penalizes wrong answers and its reported hallucination rate was lower.
How should companies use the ranking?
Use it as one comparison point, then test models on the company’s own tasks. The index score and reported benchmark-task prices do not establish which model will deliver the best accuracy, speed or cost on a specific workflow.
Source: ThorstenMeyerAI.com
