🔍 Read the full analysis: The Fine Print Behind Gemini 4 Carbon’s One-Point AI Ranking Lead on ThorstenMeyerAI.com
Get audio and creator gear 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 the Artificial Analysis Intelligence Index v5, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The ranking reflects performance across nine tests, not a universal measure of AI ability: Carbon leads four tests, loses five, and is not yet publicly available.
Google’s Gemini 4 Carbon has taken first place on the Artificial Analysis Intelligence Index v5, scoring 59 points—one ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The result makes Carbon the index leader, but the narrow margin applies to this benchmark’s nine tests, not to every kind of AI work; Google has not yet made the model publicly available.
The v5 index combines nine evaluations into a single score, with four equally weighted categories: knowledge work, coding, science and general performance. Carbon places first in four of the nine tests, including hard reasoning and knowledge questions, science workflows performed in a terminal, business workflows across software tools, and a knowledge test that penalizes wrong answers. It does not lead the other five.
The knowledge result illustrates why a ranking can conceal important differences. On AA-Omniscience, Carbon scores 54 points, six ahead of Claude Opus 5.5. But the source reports that Carbon answers 58% of questions correctly, compared with 66% for Opus 5.5 and 68% for Claude Fable 5.1. Carbon’s advantage comes from a lower reported hallucination rate: 11%, versus about 50% for Opus 5.5 and GPT-6 Astra. The scoring rewards abstaining rather than guessing when the model does not know.
Carbon also falls behind on several work categories. Claude Sonnet 5.5 and Claude Opus 5.5 lead the index’s office-style knowledge-work test; Opus 5.5 leads the real-world coding test; GPT-6 Astra leads the physics-research evaluation; Kimi K3.1 Pro leads a terminal-task test; and Meta’s Muse Wonder Full leads document reasoning. The source lists Carbon’s index-task cost at $7.76 at standard prices: less than Opus 5.5 at $10.04, but above GPT-6 Astra at $6.24 and GPT-6.1 Sol at $1.27.
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 one-point gap gives buyers a useful signal, but it is not enough on its own to identify the best model for a particular job. Carbon’s wins include science workflows and business-tool tasks, while the supplied results show other models ahead in office work, coding, physics research, terminal tasks and document reasoning. Teams using one model across different workloads may see the ordering change from task to task.
Cost also changes the practical comparison. The reported $7.76 per index task is below Opus 5.5’s stated $10.04, but above GPT-6 Astra’s $6.24 and far above GPT-6.1 Sol’s $1.27. Those figures describe the benchmark’s task costs, not necessarily the cost of every company’s workload. The source recommends testing models on a company’s own tasks and comparing the cost per completed task, rather than choosing by leaderboard position alone.
For everyday users, the immediate effect is limited: Carbon is not yet generally available. If it reaches wider release, the source expects its strengths to be most apparent on long, difficult tasks rather than routine chat. That is an expectation, not a confirmed account of real-world user results.
Products Worth Considering
DJI Osmo Pocket 3 Creator Combo Vlogging Camera with Mic 2, 1 CMOS, 4K/120fps, 3-Axis Stabilization, 128GB Card
Anker SOLIX S2000 Portable Power Station, 2,010Wh, 1,500W
EcoFlow Smart Home Panel 2 Whole-Home Backup System Hub
As an affiliate, we earn on qualifying purchases.
How the V5 Index Builds Its Score
Artificial Analysis introduced version 5 on October 27, 2026. It groups the nine tests into four blocks, each worth 25% of the overall result. 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; and the general category includes AA-Omniscience v1.1 and HLE-Diamond.
According to the supplied source, 60% of v5 uses private test sets or held-out answers, reducing the chance that a model can rely on memorized questions. Many evaluations involve extended agent tasks, such as working in a terminal, producing reports or modifying code. That design tests more than short question-and-answer exchanges, but also makes the evaluations more expensive to run.
The score is therefore a comparison at particular model settings against a particular set of tests. The source reports Carbon in a high-reasoning configuration, Opus 5.5 and GPT-6 Astra at maximum settings, and other models at their listed settings. A change in the evaluations, model configuration or intended task can produce a different ordering. The index is not a standalone verdict on which system is “the smartest.”
“A subset of cyber defenders”
— Google, according to the supplied source
Access, Pricing and Real-World Results
Public availability remains undated. The supplied source says Google plans to provide Carbon first to a subset of cyber defenders, then to trusted partners, followed by developers, businesses and consumers. It does not give a calendar date for any of those stages, or explain how the initial group will be selected.
The source also gives introductory API prices of $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20 respectively. It does not specify when the introductory rates end, what conditions apply, or how those token prices map to the per-task benchmark cost. That makes direct price comparisons dependent on workload and usage.
The material does not provide enough detail to independently assess all test procedures or reproduce the full ranking. It also does not establish how Carbon performs on customers’ own tasks, how its costs will vary in deployment, or whether the benchmark lead will hold as tests and model versions change. The reported one-point margin should not be treated as proof of a broad performance advantage.
Google’s Staged Carbon Rollout
The next development to watch is Google’s planned access sequence: an initial group of cyber defenders, followed by trusted partners, developers, businesses and consumers. No rollout dates are supplied, so users and organizations cannot yet plan around a confirmed public launch.
When access becomes available, buyers can compare Carbon against alternatives on their own work, including output quality, reliability, task completion time and total cost. The index provides a starting point for those evaluations, but the actual choice will depend on what an organization needs the model to do and what it costs to complete that work.
Key Questions
How far ahead is Gemini 4 Carbon?
Carbon scores 59 points on the Artificial Analysis Intelligence Index v5, one point ahead of Claude Opus 5.5 at 58 and two ahead of GPT-6 Astra at 57.
Does Carbon lead all nine tests?
No. The supplied results say Carbon leads four of nine tests and loses the other five. Its competitors lead in several coding, office-work, science and document-reasoning evaluations.
Why does Carbon win the knowledge test if it answers fewer questions correctly?
AA-Omniscience penalizes wrong answers. Carbon’s reported accuracy is lower than that of some rivals, but its 11% hallucination rate is also lower than the roughly 50% reported for Claude Opus 5.5 and GPT-6 Astra. The scoring rewards avoiding unsupported guesses.
Can the public use Gemini 4 Carbon now?
Not according to the supplied source. Google is starting with a subset of cyber defenders, then plans to expand access through partners and other groups. No public-release date is given.
Which model is cheapest in the reported benchmark costs?
The source lists GPT-6.1 Sol at $1.27 per index task, compared with $6.24 for GPT-6 Astra, $7.76 for Carbon and $10.04 for Claude Opus 5.5. These are reported benchmark task costs, not a guarantee of costs on other workloads.
Source: ThorstenMeyerAI.com
