🔍 Read the full analysis: Gemini 4 Carbon Is #1 By A Point—What That AI Ranking Really Tells You 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 leads Artificial Analysis’s Intelligence Index v5 with 59 points, one ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The result reflects performance across nine tests, not a universal measure of AI capability; Carbon is not publicly available, and its cost and uneven task results matter.
Gemini 4 Carbon has moved to the top of Artificial Analysis’s Intelligence Index v5, scoring 59 points—one ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The narrow lead is a result on one nine-test benchmark, not a general finding that Carbon is the best model for every task; Google has also not made it publicly available.
Artificial Analysis rebuilt its leaderboard and launched version 5 on October 27. The index gives equal weight to four areas: knowledge work, coding, science and general capability. It combines results from nine tests, including business workflows, coding tasks, scientific work in a terminal and questions that reward models for not guessing when uncertain.
Carbon, evaluated at its high-reasoning setting, placed first in four of the nine tests, according to the source report: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. It did not lead the other five. Claude models scored higher on office-style knowledge work, Claude Opus 5.5 led on real-world coding, GPT-6 Astra led on a physics research test, Kimi K3.1 Pro led on terminal tasks, and Meta’s Muse Wonder Full led on document reasoning.
The report puts Carbon’s cost 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 GPT-6.1 Sol at $1.27. Google’s introductory API pricing is listed as $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20, respectively. These are reported prices; actual costs depend on usage and task setup.
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 ranking can help readers compare models on a shared set of tests, but its one-point margin is small and its outcome depends on the benchmark’s choices. A company choosing a model for coding, document analysis or office work may get a different answer from the overall leaderboard. The test-level results show why the aggregate score should not replace task-specific evaluation.
Price also changes the practical comparison. Carbon costs less per index task than Claude Opus 5.5 in the cited figures, but more than GPT-6 Astra and substantially more than GPT-6.1 Sol. For businesses, the relevant measure is not rank alone but cost per successful task, including how often a model needs retries, human review or correction.
The report says Carbon’s result on AA-Omniscience is driven partly by its willingness to abstain rather than answer uncertain questions. It scores 58% on plain accuracy, below several Claude and GPT models, but has an 11% hallucination rate on that test, compared with about 50% for Claude Opus 5.5 and GPT-6 Astra. That distinction matters in settings where an unsupported answer is more costly than an admitted gap in knowledge. It does not, by itself, establish how often Carbon will make errors in everyday use.
Products Worth Considering
Samsung 27-Inch Odyssey QD-OLED G8 4K 240Hz Gaming Monitor
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
As an affiliate, we earn on qualifying purchases.
How the New Index Is Built
Artificial Analysis describes its Intelligence Index as a combined score from nine tests, split into four equally weighted blocks. 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 the hard subset of CritPt v2; general capability includes AA-Omniscience and HLE-Diamond.
According to the source report, 60% of the index now uses private test sets or held-out answers, reducing the scope for models to rely on memorized questions. Many tasks also involve longer agent workflows, such as working in a terminal, producing reports or modifying code. That makes the benchmark different from a simple set of short question-and-answer prompts, but does not make it a complete measure of performance across all products or user needs.
The leaderboard’s top reported scores are 59 for Gemini 4 Carbon, 58 for Claude Opus 5.5, 57 for GPT-6 Astra and 55 for GPT-6.1 Sol. Google’s previous model, Gemini 4 Argon, scored 50. Those results compare particular models at particular settings; changes to settings, tests or weighting could produce a different ordering.
““Choose by task and by cost per finished task, not by rank.””
— Thorsten Meyer, source report
Access and Real-World Performance
Google’s rollout schedule has no stated dates. The source says access begins with a subset of cyber defenders, followed by trusted partners, then developers, businesses and consumers. It does not specify how large each group will be, when public access might begin, or whether the sequence could change.
The supplied reporting also does not establish how closely benchmark results will match performance on a particular company’s software, documents or coding environment. The index’s cost figure is tied to its own tasks and stated standard prices; it may not predict a customer’s total cost. Carbon’s reported hallucination rate applies to AA-Omniscience, not every type of use.
The source excerpt notes that Carbon loses several tests, but ends before giving full results for all of them. The ranking and the test-level figures here should therefore be read as the reported snapshot, not a complete account of every comparison or subsequent change.
Google’s Staged Rollout
Google’s stated next step is a limited rollout to cyber defenders, then access for trusted partners, developers, businesses and eventually consumers. No timetable was supplied. Until access expands, most users cannot test Carbon directly or judge whether its benchmark strengths hold on their own work.
For organizations considering a switch, the next useful step is a controlled comparison on representative tasks: measure accuracy, completion time, required human review and total cost. The headline score may guide which models to try, but the index does not settle which one is best for a specific workflow.
Key Questions
What does Gemini 4 Carbon’s 59-point score mean?
It is Carbon’s score on Artificial Analysis’s Intelligence Index v5, which combines results from nine tests across knowledge work, coding, science and general capability. It does not measure every use of an AI model.
How far ahead is Carbon?
Carbon scored 59 points, one more than Claude Opus 5.5 at 58 and two more than GPT-6 Astra at 57. The result is a narrow lead on this index.
Can the public use Gemini 4 Carbon now?
No public availability date is given in the source report. Google is reported to be starting with a subset of cyber defenders, with wider access planned later.
Is Carbon the strongest model for coding?
Not according to the test results described in the report. Claude Opus 5.5 led Carbon on the real-world coding test cited, and Carbon did not top all nine index tests.
Does the index show which model is best for a business?
No. It offers one comparison across selected tests. Businesses should evaluate models on their own tasks and compare quality, human review needs and cost, rather than relying on the overall rank alone.
Source: ThorstenMeyerAI.com
