🔍 Read the full analysis: Gemini 4 Carbon Is The New #1 AI Model By One Point—But What’s Missing? 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 scored 59 on Artificial Analysis’ Intelligence Index v5, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The ranking covers nine tests, and Carbon leads on four; Google has not made it publicly available, beginning instead with a limited rollout to cyber defenders.
Google’s Gemini 4 Carbon has taken the top position in the Artificial Analysis Intelligence Index v5, scoring 59 points—one ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The narrow lead is based on nine benchmark tests, not a general measure of every AI capability, and Carbon is not yet publicly available.
The index’s published scores put Claude Opus 5.5 at 58, GPT-6 Astra at 57, GPT-6.1 Sol at 55 and Google’s earlier Gemini 4 Argon at 50. Carbon ranks first in four of the nine tests, according to the source report: hard reasoning and knowledge questions, science workflows in a terminal, knowledge questions that penalize incorrect answers, and business workflows. It does not lead the other five.
Its strongest individual results include 69% on HLE-Diamond, against 68% for Claude Opus 5.5, and 65% on Terminal-Bench-Science, against 63% for GPT-6 Astra. On AutomationBench-AA, Carbon scores 78%, just ahead of its predecessor, Gemini 4 Argon. These results reflect specific benchmark tasks and should not be read as evidence that Carbon leads every kind of work.
The ranking has a cost and speed dimension too. The source report puts Carbon’s estimated cost at $7.76 per index task, compared with $10.04 for Claude Opus 5.5 and $6.24 for GPT-6 Astra. It also characterizes Carbon as slow on long tasks. Google’s stated introductory API price is $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20 respectively; the model is currently being offered to a limited group of cyber defenders, with no public release date given.
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
A one-point difference makes the result a useful signal, not a decisive verdict. The index combines performance across a defined set of tests, and the score does not show whether the difference is meaningful for a particular company’s workload. Carbon’s mixed results make that distinction concrete: it leads some reasoning, science and workflow tasks but trails competitors in office-style knowledge work and coding, according to the report.
For organizations, the practical question is not simply which model ranks first. It is whether the model completes their own tasks accurately, quickly and at an acceptable cost per finished job. An expensive model that uses fewer retries may suit one workflow; a cheaper model may be preferable for routine work. Those comparisons require testing on the organization’s own data and tools.
For individual users, there is no immediate product change because Carbon is not publicly accessible. If and when broader access arrives, the benchmark suggests its advantage may be most relevant on demanding, multi-step tasks rather than everyday chat. That is an interpretation of the test profile, not a confirmed description of user experience.
Popular right now
Dyson Cyclone V10 Origin Cordless Vacuum
eufy E25 Robot Vacuum and Mop Combo with HydroJet Roller Mop
Samsung 48-Inch S90H OLED 4K Smart TV (2026)
As an affiliate, we earn on qualifying purchases.
What the New Index Measures
Artificial Analysis launched version 5 on October 27, rebuilding its leaderboard around nine tests grouped into four equally weighted areas: knowledge work, coding, science and general performance. The source report says 60% of the tests use private test sets or held-out answers, intended to reduce the chance that models have memorized the questions. Many tasks involve extended agent work, such as operating in a terminal, preparing reports or changing code.
The index is therefore best understood as a combined score for one model configuration on these nine tests. The report says Carbon leads HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. It also says Claude models lead office-style knowledge work, Claude Opus 5.5 leads the cited real-world coding test, GPT-6 Astra leads a physics research test, and Kimi K3.1 Pro and Meta’s Muse Wonder Full lead other listed categories. Different tests or weightings could produce a different order.
One result highlights how scoring rules shape rankings. Carbon scores 58% accuracy on AA-Omniscience, below several listed Claude and GPT models, but receives 54 points on the test because incorrect answers are penalized. Artificial Analysis reports an 11% hallucination rate for Carbon, versus about 50% for Claude Opus 5.5 and GPT-6 Astra under that test’s definition. The figures describe this benchmark, not every response those models produce.
Limits Behind the Leaderboard
The source material does not provide the full benchmark tables, test methodology or underlying data for independent review here. It describes the index’s categories and several results, but readers cannot determine from the supplied account alone how much uncertainty surrounds the one-point gap or whether that margin would hold under a different test mix.
Access remains another open question. Google has described an initial rollout to a subset of cyber defenders, followed by trusted partners, developers, businesses and consumers, but has not provided dates for those stages. The source also says Carbon thinks slowly, but does not supply comparable latency measurements. The quoted passage about Artificial Analysis’ coding-agent comparison is incomplete, so no further conclusion about Google’s tool performance can be drawn from it.
Access and Independent Testing
The next reported step is Google’s limited release to selected cyber defenders, followed by planned access for trusted partners and, later, developers, businesses and consumers. No timetable has been announced, so public availability remains uncertain. The introductory and later API prices reported by the source may also matter as businesses evaluate whether to test the model.
As access expands, the most useful comparisons will be direct tests of the same tasks, tools, settings and pricing across competing models. Until then, the leaderboard establishes that Carbon leads this version of this index by one point; it does not establish that it is the best choice for every user or workload.
Key Questions
What does Gemini 4 Carbon’s number-one ranking mean?
It means Carbon scored 59 points on Artificial Analysis Intelligence Index v5, one point above Claude Opus 5.5. The result applies to the index’s nine tests and their weighting, not to every AI task.
Is Gemini 4 Carbon available to the public?
No. The source report says Google is starting with a subset of cyber defenders, then plans to expand access. It gives no date for public availability.
Where does Carbon perform best?
According to the report, Carbon ranks first on four tests: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. Its position varies across the other tests.
Does Carbon know more than the other leading models?
Not on the plain-accuracy measure cited for AA-Omniscience. Several Claude and GPT models answered more questions correctly, while Carbon scored well on the version that penalizes wrong answers and reports a lower hallucination rate.
How much does Carbon cost to run?
The report estimates $7.76 per index task. It also lists introductory API prices of $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20. Actual costs depend on use.
Source: ThorstenMeyerAI.com
