🔍 Read the full analysis: Gemini 4 Carbon Ranks #1 By One Point—How To Make Sense Of The Result on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
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 index lead reflects performance across nine selected tests, not a universal ranking; Carbon wins four tests and loses five, and Google has not made it publicly available.
Google’s Gemini 4 Carbon ranks first on the Artificial Analysis Intelligence Index v5 with 59 points, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The narrow result puts Carbon atop this particular nine-test benchmark, but does not establish it as the best model for every task—and Google has not opened it to the public.
The index, rebuilt by Artificial Analysis and launched on October 27, combines results from nine tests in four equally weighted categories: knowledge work, coding, science and general performance. Carbon is listed in its high-reasoning configuration. The reported scores are 59 for Carbon, 58 for Claude Opus 5.5 at maximum reasoning, 57 for GPT-6 Astra at maximum reasoning, 55 for GPT-6.1 Sol at maximum reasoning and 50 for Google’s previous model, Gemini 4 Argon.
Carbon places first on four tests: HLE-Diamond, a difficult reasoning and knowledge exam; Terminal-Bench-Science; AA-Omniscience; and AutomationBench-AA, which tests business workflows across software tools. It does not lead the other five. Claude models score higher on office-style knowledge work, Claude Opus 5.5 leads the cited real-world coding test, GPT-6 Astra leads the physics research test, Kimi K3.1 Pro leads terminal tasks, and Meta’s Muse Wonder Full leads document reasoning.
The index source reports a cost of $7.76 per index task for Carbon at standard prices. That is below the stated $10.04 for Claude Opus 5.5, but above GPT-6 Astra’s $6.24 and GPT-6.1 Sol’s $1.27. Separately, Google’s introductory API price is listed as $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20 respectively. These are different pricing measures: the task-cost figure refers to the benchmark run, while token prices describe API usage.
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
For companies choosing an AI model, the one-point gap is a signal to compare performance on their own work, not a reason to switch automatically. A model’s overall score combines categories that may not match a particular organization’s needs. Carbon’s reported wins in science workflows, hard questions and business automation could matter to users doing those tasks; its lower result in office-style knowledge work or coding may matter more to others.
Price also depends on what a model completes, not only its listed rate or leaderboard position. The reported $7.76 benchmark-task cost sits between the prices of several close competitors, and the source says Carbon’s responses can be slow. Organizations should weigh task quality, completion time, model setting and cost together. The ranking offers a comparison point, but it does not supply a universal measure of value.
The result is also relevant to individual users mainly as an indication of possible future capability. There is no public access today, and the source characterizes Carbon’s advantages as most visible on long, difficult tasks rather than ordinary chat. Its immediate effect on consumer AI use is therefore limited.
Popular right now
De'Longhi Magnifica Evo Fully Automatic Espresso Machine with Manual Milk Frother, Silver
De'Longhi 15-Bar Pump Espresso Machine with Milk Frother & Temperature Control
Roborock F25 GT Set Cordless Wet Dry Vacuum Cleaner and Mop with 2 Brushes
As an affiliate, we earn on qualifying purchases.
How the Nine-Test Index Works
Artificial Analysis describes version 5 as a combined score across four equally weighted blocks. Knowledge work accounts for 25% of the index through AA-Briefcase, AutomationBench-AA and GDP.pdf. Coding contributes 25% through FrontierCode 2.0 and Terminal-Bench 4.0; science contributes 25% through Terminal-Bench-Science and the hard subset of CritPt v2; and general performance contributes 25% through AA-Omniscience and HLE-Diamond.
The source says 60% of v5 uses private test sets or held-out answers, intended to reduce the chance that models have simply memorized test material. Many tasks require extended agent work, such as using a terminal, producing reports or changing code. That makes the benchmark more task-oriented than a simple set of short questions, but also means its outcome depends on the test selection, model configuration and evaluation method.
One result illustrates why individual test definitions matter. Carbon scores 58% on plain accuracy in AA-Omniscience, behind several Claude and GPT models. Yet it leads that test’s score because the measure penalizes wrong answers and rewards restraint. Artificial Analysis reports an 11% hallucination rate for Carbon on the measure, compared with about 50% for Claude Opus 5.5 and GPT-6 Astra. Those figures describe this benchmark’s scoring setup; they should not be read as general error rates across all uses.
Access and Real-World Performance
Several points remain unsettled. Google’s staged release is described as starting with a subset of cyber defenders, followed by trusted partners, developers, businesses and consumers, but the source gives no public release date. It also does not specify how many testers will receive access or what criteria will govern each stage.
The supplied material gives a summary of the index and selected results, but not the full benchmark report, testing conditions or uncertainty ranges around scores. The one-point difference should be understood as the reported leaderboard margin; the available information does not establish whether it is statistically meaningful or likely to hold under different settings. Nor does a score across nine tests settle how Carbon will perform on a company’s own data, software or workflows.
Some comparisons use different model configurations, including maximum reasoning settings for the listed Claude and GPT models. The source also cites a separate coding-agent comparison that combines a model with its coding tool. Those results are not the same as a model-only score, and the source excerpt cuts off during its discussion of that comparison. Further detail on methods and limitations is needed before drawing broader conclusions.
Google’s Staged Release Plan
According to the source material, Google plans to provide Carbon first to a subset of cyber defenders, then to trusted partners, with broader access for developers, businesses and consumers planned afterward. The company has not attached dates to those stages in the information provided. Access announcements and published API terms will clarify when people outside the initial groups can test it and how the listed introductory prices apply.
For now, readers and organizations can treat the index as a benchmark snapshot dated October 2026. A useful next check is whether independent testing, fuller methodological disclosures and real-world evaluations support the same task-specific strengths. For buyers, the practical next step—once access is available—is to test Carbon against alternatives on representative work and compare completed-task quality, speed and cost, rather than choosing by rank alone.
Key Questions
What score did Gemini 4 Carbon receive?
It received 59 points on Artificial Analysis’ Intelligence Index v5, according to the supplied source. Claude Opus 5.5 scored 58 and GPT-6 Astra scored 57.
Does the ranking show that Carbon is the best AI model?
No. It shows that Carbon ranked first on this nine-test index. It led four tests and lost five, so its standing can differ by task and by the tests a user values.
Can the public use Gemini 4 Carbon now?
No public availability is reported. Google is said to be starting with a subset of cyber defenders, then expanding to trusted partners and later groups. No release date is given.
Why did Carbon lead the AA-Omniscience test if other models answered more correctly?
The test’s score penalizes wrong answers, not just missed questions. Carbon’s reported 58% accuracy was lower than several rivals, but its lower reported hallucination rate helped it score highest under that test’s rules.
How much does Carbon cost?
The source reports $7.76 per index task at standard prices and an introductory API rate of $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20. The task-cost and token-price figures measure different things.
Source: ThorstenMeyerAI.com
