🔍 Read the full analysis: Why Gemini 4 Carbon’s New #1 AI Model Ranking Needs More Context on ThorstenMeyerAI.com
Get hardware and tech essentials 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’s Intelligence Index v5, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. It led on four of nine tests, while losing on five; the index score does not establish a universal best model. Google says Carbon is initially limited to a subset of cyber defenders, with no public release date announced.
Google’s Gemini 4 Carbon has taken the top spot in Artificial Analysis’s Intelligence Index v5, scoring 59 points—one ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The result is a narrow lead on one benchmark, not a finding that Carbon is the best model for every task; it currently leads on four of nine tests and is not publicly available.
Artificial Analysis rebuilt its leaderboard and launched index version 5 on October 27, according to the source. Carbon’s high-reasoning configuration scored 59, compared with 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. These are index scores, not percentages or direct measures of performance on every real-world task.
Carbon placed first on HLE-Diamond, a hard reasoning and knowledge test; Terminal-Bench-Science; AA-Omniscience; and AutomationBench-AA. It did not lead on the other five tests. Claude models scored higher on office-style knowledge work, Claude Opus 5.5 led on the cited real-world coding benchmark, GPT-6 Astra led on physics research, Kimi K3.1 Pro on terminal tasks, and Meta’s Muse Wonder Full on document reasoning.
The source puts the index’s estimated cost at $7.76 per task at standard prices. That is below the stated $10.04 for Claude Opus 5.5 at maximum reasoning, but above GPT-6 Astra’s $6.24 and GPT-6.1 Sol’s $1.27. It also reports an introductory API price of $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20 respectively. The source does not provide further pricing conditions here.
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 matters because it offers a common comparison across reasoning, coding, science and knowledge-work tasks. But the one-point gap between Carbon and Claude Opus 5.5 is small, and the underlying results show different models leading on different work. A company choosing a model based only on the overall position could miss a stronger or cheaper option for its own tasks.
Cost and availability also limit the score’s immediate practical impact. The reported per-task estimate places Carbon above two named rivals, and Google has not made it broadly available. For organizations, the more useful comparison is likely to be performance and cost on their own workloads, including the cost of completing a task rather than the model’s headline rank alone. The source recommends testing on company-specific work before switching.
The result may be more relevant to people handling long, demanding tasks than to everyday chat users, according to the source. It does not establish that ordinary users will see an immediate change: Carbon is not yet generally accessible, and the source gives no consumer launch date.
Products Worth Considering
Logitech MX Keys S Wireless Keyboard, Graphite
Apple MacBook Pro 16-inch (2021) – M1 Pro Chip, 10-Core CPU, 16-Core GPU, 16GB RAM, 1TB SSD, Space Gray
Ledger Flex Crypto Wallet
As an affiliate, we earn on qualifying purchases.
How the Nine-Test Index Works
Artificial Analysis describes v5 as a combined score across nine tests, divided into four equally weighted blocks: knowledge work, coding, science and general. The tests include agent-style work in terminals, business workflows using software tools, reports and code tasks. The index therefore measures performance on this specific selection of tasks and configurations; changing the tests or model settings could change the order.
The source says 60% of the test material uses private test sets or held-out answers, intended to reduce the chance that models have memorized questions. It also notes that many evaluations involve longer agent tasks, which it says are more realistic but more expensive to run. The index is not a general verdict on which model is “the smartest.”
One of Carbon’s wins illustrates how a benchmark’s scoring rules shape the result. On AA-Omniscience, Carbon scored 54 points, ahead of Claude Opus 5.5 by six. Yet the source reports lower plain-answer accuracy for Carbon—58%—than for Claude Fable 5.1 at 68%, Claude Opus 5.5 at 66%, or GPT-6.1 Sol and GPT-6 Astra at 62%. The benchmark penalizes wrong answers, and Carbon’s reported hallucination rate was 11%, compared with about 50% for Claude Opus 5.5 and GPT-6 Astra. That rewards caution, not simply the largest number of correct answers.
““Choose by task and by cost per finished task, not by rank.””
— Thorsten Meyer, source article
Release Timing and Benchmark Limits
Google has not provided a public release date in the supplied material. The stated rollout begins with a subset of cyber defenders, then trusted partners, followed by developers, businesses and consumers. The timing and criteria for each stage remain unspecified.
The available source excerpt also ends mid-sentence while discussing Artificial Analysis’s comments on a separate coding-agent comparison. It reports scores of 62 for Claude Code with Claude Opus 5.5 and 59 for Google’s Antigravity CLI with Carbon, but does not include the full accompanying explanation. Those figures should not be treated as a complete account of the comparison.
More broadly, the supplied material does not establish how Carbon will perform across every deployment, prompt, tool setup or customer workload. Its rank reflects the nine tests and settings used in v5. The score alone does not resolve whether its cost, speed or error profile will suit a particular user; the source describes Carbon as slow-thinking but gives no response-time measurements.
Google’s Staged Rollout
According to the source, Google plans to make Carbon available first to a subset of cyber defenders, then to trusted partners, developers, businesses and consumers. No dates are supplied for those stages, so the pace and eventual public availability remain open.
Further comparisons may clarify how Carbon performs outside the index, particularly on the tasks where it did not lead and on the cost of completed work. Until access broadens, the ranking is a benchmark result rather than a product most users can try. Organizations considering it will need to compare it with alternatives on their own tasks and account for both output quality and usage costs.
Key Questions
Did Gemini 4 Carbon decisively beat the other models?
No. It scored 59 points, one ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The result is a narrow lead in Artificial Analysis’s v5 index, not a win on every test.
What does the index measure?
It combines results from nine tests grouped into knowledge work, coding, science and general tasks. The source describes tests involving agent workflows, software tools, coding and difficult reasoning questions.
Is Gemini 4 Carbon available to the public?
Not according to the supplied material. Google is reported to be starting with a subset of cyber defenders, followed by other groups, but no public release date is given.
Why did Carbon lead AA-Omniscience if other models answered more correctly?
The test penalizes wrong answers as well as rewarding correct ones. The source reports that Carbon answered fewer questions correctly than several rivals but had a lower reported hallucination rate, which helped it score higher under that test’s rules.
What should companies take from the ranking?
The index is one comparison, not a substitute for testing a model on a company’s own work. The source advises comparing task performance and cost per completed task before making a switch.
Source: ThorstenMeyerAI.com
