🔍 Read the full analysis: One Point Ahead: Gemini 4 Carbon’s Place In The AI Model Rankings 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 four of nine tests, but the ranking measures a specific set of tasks, Carbon is costly to run, and Google has not made it publicly available.
Google’s Gemini 4 Carbon, running in high-reasoning mode, scored 59 points on Artificial Analysis’ Intelligence Index v5, edging Claude Opus 5.5 by one point and GPT-6 Astra by two. The result puts Carbon at the top of this particular nine-test ranking, but the model is not yet publicly available, and its performance varies across the tests that make up the score.
The leaderboard lists Claude Opus 5.5 at 58, GPT-6 Astra at 57 and GPT-6.1 Sol at 55. Google’s previous model, Gemini 4 Argon, scored 50. Carbon placed first in four of the nine tests: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. The supplied results show it narrowly ahead on some measures: 69% to Claude Opus 5.5’s 68% on HLE-Diamond, and 65% to GPT-6 Astra’s 63% on Terminal-Bench-Science.
Carbon’s lead does not extend across the full test set. Claude models score higher on office-style knowledge work, Claude Opus 5.5 leads on real-world coding, GPT-6 Astra leads on physics research, Kimi K3.1 Pro leads on terminal tasks, and Meta’s Muse Wonder Full leads on document reasoning. On AA-Briefcase, Carbon scored 1,609 Elo, below Claude Sonnet 5.5 at 1,812 and Claude Opus 5.5 at 1,800. On FrontierCode 2.0, it scored 41%, compared with 46% for Opus 5.5.
Cost and access are practical constraints. Artificial Analysis estimates Carbon costs $7.76 per index task at standard prices: less than Opus 5.5’s $10.04, but more than GPT-6 Astra’s $6.24 and well above GPT-6.1 Sol’s $1.27. Google says it is first giving access to a subset of cyber defenders, followed by trusted partners and then developers, businesses and consumers. No public release date was provided.
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 may matter to organizations evaluating models for long, multi-step work, including research, tool use and business workflows. Carbon’s four test wins indicate strengths in parts of that workload, while its weaker results on coding and office-style tasks caution against treating the overall score as a universal measure of usefulness. A company choosing a model for a particular job may get a different answer from the leaderboard’s top line.
The benchmark’s cost estimate also makes a ranking point an incomplete buying guide. The relevant comparison for businesses is the cost per successfully completed task, not just cost per benchmark run or a model’s overall position. The figures supplied do not establish what Carbon will cost across every real-world workload, and performance may depend on the tools, prompts and settings used.
For ordinary users, there is no immediate product change: Carbon is not generally available. If Google’s staged rollout reaches consumers, the source suggests its strengths may be most apparent on difficult, extended tasks rather than routine chat. That remains a forecast, not a public-user finding.
Products Worth Considering
Jackery HomePower 3000 Portable Power Station with Two 200W Solar Panels
Apple AirPods Pro 3 Wireless Earbuds
Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Black
As an affiliate, we earn on qualifying purchases.
What Intelligence Index v5 Measures
Artificial Analysis launched Index v5 on October 27, organizing its nine tests into four equally weighted groups: knowledge work, coding, science and general capability. The individual tests cover tasks such as creating reports, using software tools, writing or repairing code, conducting science workflows, and answering difficult knowledge questions. Models are assessed at particular settings, so Carbon’s score refers to its high-reasoning configuration rather than every possible version of the model.
According to the supplied account of the benchmark, 60% of the tests use private test sets or held-out answers, intended to reduce the benefit of memorizing public questions. Many tasks are lengthy and involve an AI agent working through a terminal or other tools. That design aims to reflect more complex work, but it also means the index remains a defined selection of tasks rather than a general verdict on intelligence or performance in every setting.
One result highlights how a test’s scoring rules can change the ranking. Carbon scored 58% on plain accuracy in AA-Omniscience, below Claude Fable 5.1 at 68%, Claude Opus 5.5 at 66%, and GPT-6.1 Sol and GPT-6 Astra at 62%. Yet Carbon topped the test’s penalized score with 54 points. Artificial Analysis reports an 11% hallucination rate for Carbon, versus about 50% for Opus 5.5 and Astra, under the benchmark’s definition. The result rewards a model for abstaining rather than giving an incorrect answer; it does not mean Carbon answered the most questions correctly.
Public Access and Real-World Results
Google has not announced a public release date or a timetable for moving from cyber defenders to partners, developers, businesses and consumers. The reported API pricing is introductory: $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20, respectively. The source does not specify when those prices change or how broadly they will apply.
The supplied material also does not provide enough detail to independently evaluate all nine test results, the benchmark’s full scoring procedures or the reliability of the cost estimates across different workloads. Index placement can shift with test selection, model settings and tool configuration. Nor does the ranking establish that Carbon will outperform rivals in every company’s own tasks. These limits make direct testing important before a purchase or migration decision.
Google’s Staged Rollout
Google’s stated plan is to give Carbon to a subset of cyber defenders first, then trusted partners, followed by developers, businesses and consumers. No dates were supplied for those steps. Until access broadens, the index result is more relevant to benchmark comparisons than to users seeking to try the model.
For organizations, the next useful step is to test Carbon against competing models on their own representative tasks, tracking quality, completion rates, latency and total cost. Artificial Analysis’ score is a starting comparison, not a substitute for those evaluations. Further information about availability, standard API prices and performance outside the benchmark remains pending.
Key Questions
Did Gemini 4 Carbon win every test?
No. It placed first on four of nine tests. The supplied results say rivals lead in office-style work, real-world coding, physics research, terminal tasks and document reasoning.
What does a score of 59 mean?
It is Carbon’s overall score on Artificial Analysis’ Intelligence Index v5, which combines nine tests grouped into four equally weighted areas. It is not a universal measure of intelligence or a guarantee of performance on a particular task.
Is Gemini 4 Carbon available to the public?
Not yet, according to the supplied report. Google is starting with a subset of cyber defenders, then plans to expand access, but has not given dates for the stages.
Why did Carbon lead AA-Omniscience if other models had higher accuracy?
The test penalizes incorrect answers, and Carbon reportedly had a lower hallucination rate. Its 58% plain accuracy was below several rivals, but it scored highest under the test’s penalty-based measure.
Is Carbon the cheapest model in the top group?
No. The reported estimate is $7.76 per index task, compared with $6.24 for GPT-6 Astra and $1.27 for GPT-6.1 Sol. Costs can vary by task and usage, and the source does not establish a universal real-world cost comparison.
Source: ThorstenMeyerAI.com
