🔍 Read the full analysis: Look Beyond The Score: Gemini 4 Carbon’s One-Point AI Model Lead on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- 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 result reflects performance on nine specific benchmarks, not an across-the-board ranking: Carbon leads four tests, trails on five and is not yet publicly available.
Google’s Gemini 4 Carbon has taken the top position on the Artificial Analysis Intelligence Index v5, scoring 59 points, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The benchmark result is a narrow lead across nine tests, not evidence that Carbon outperforms competitors on every task; Google has begun with a limited rollout to cyber defenders, and the model is not yet publicly available.
Artificial Analysis, an independent benchmarking company, introduced version 5 of its index on October 27. Its reported scores put Claude Opus 5.5 at 58, GPT-6 Astra at 57, GPT-6.1 Sol at 55 and Google’s previous model, Gemini 4 Argon, at 50. The index combines nine tests in four equally weighted groups: knowledge work, coding, science and general performance.
Carbon ranks first on four of the nine tests: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. It does not lead the other five. Claude models score higher on office-style knowledge work, Claude Opus 5.5 leads the real-world coding test, GPT-6 Astra leads the physics research benchmark, and Kimi K3.1 Pro and Meta’s Muse Wonder Full top other tasks named in the source report.
The result also depends on how a test rewards uncertainty. On AA-Omniscience, Carbon scores 54 points, while the report gives its plain-answer accuracy as 58%, below several Claude and GPT models. Artificial Analysis’s reported hallucination rate for Carbon is 11%, compared with about 50% for Claude Opus 5.5 and GPT-6 Astra. The test penalizes wrong answers, so a model that declines to guess can place ahead even if it answers fewer questions correctly.
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 offers companies a useful comparison, but its one-point margin is small and tied to a defined set of tasks and model settings. The index itself says more about performance on its selected benchmarks than about how a model will perform across every company’s workflows, software or prompts. A top-line score should not replace task-specific testing.
Cost and speed also affect practical value. The source report estimates $7.76 per index task at standard prices. That is lower than its estimate of $10.04 for Claude Opus 5.5 at maximum settings, but higher than $6.24 for GPT-6 Astra and $1.27 for GPT-6.1 Sol. These are reported benchmark-task costs, not a universal bill for every use. The report also describes Carbon as slow to produce answers, though it does not provide a response-time figure in the supplied material.
For everyday users, the immediate impact is limited because access remains restricted. If the ranking holds up on other evaluations, Carbon’s strongest appeal may be for demanding reasoning, science and business-process tasks rather than ordinary chat. Whether that advantage offsets cost, latency and access limits will depend on the work being done.
Products Worth Considering
HP Sprocket Photo Booth Machine Instant Color Photo Printer with 10.1 Touchscreen, LED Ring Light, 250+ Frames, 3×4 ZINK Paper, QR Code Shar
Dell UltraSharp U2725QE 27-Inch 4K UHD IPS Black Monitor
FLEXISPOT EN1 Electric Standing Desk, 48 x 24 in, Black
As an affiliate, we earn on qualifying purchases.
How Version 5 Builds Its Score
Artificial Analysis describes its index as a combined score across nine evaluations, grouped into four blocks with equal weight: knowledge work, coding, science and general performance. The source report says 60% of version 5 uses private test sets or held-out answers, intended to reduce the chance that models have memorized questions. Many evaluations involve extended agent tasks, such as working in a terminal, producing reports or changing code, rather than answering short chat prompts.
The tests assign different weights within each block. For example, FrontierCode 2.0 accounts for 15% of the full index, as does AA-Briefcase and Terminal-Bench-Science. HLE-Diamond contributes 10%. The resulting total compresses different kinds of capability into one number; shifting the tests or their weights could change the order. The source article says Carbon scores 69% on HLE-Diamond, against 68% for Claude Opus 5.5, and 65% on Terminal-Bench-Science, compared with 63% for GPT-6 Astra.
Availability is a separate issue from the benchmark result. According to the source, Google is first providing Carbon to a subset of cyber defenders, then plans access for trusted partners, developers, businesses and consumers. It gives no timetable for those stages. The report lists introductory API pricing of $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20 respectively; the timing and full terms were not specified in the supplied material.
Release Timing and Real-World Performance
No public release date is given, and the supplied report does not specify how long each stage of Google’s rollout will take. It also does not provide enough detail to independently verify the benchmark runs, the cost assumptions behind the per-task estimates or Carbon’s latency relative to competitors.
The index score cannot establish how Carbon will perform on a particular organization’s data, tools or workload. The supplied material also ends during its discussion of the separate Coding Agent Index, so it does not include the full explanation of the comparison between Google’s Antigravity CLI and Claude Code. The reported results should be treated as benchmark findings, not as a guarantee of performance or savings in deployment.
Google’s Staged Rollout and Testing
Google’s stated sequence is to extend access from selected cyber defenders to trusted partners, then to developers, businesses and consumers. No dates for those steps have been announced in the supplied source. Broader availability and published product details would clarify which interfaces, plans and usage limits users can expect.
For organizations weighing a switch, the next useful evidence is performance on their own tasks, alongside actual cost per completed task and acceptable response times. Further independent evaluations could show whether Carbon’s lead persists beyond the nine benchmarks in version 5. Until then, the confirmed development is a narrow leaderboard advantage, paired with restricted access rather than a general public launch.
Key Questions
What score did Gemini 4 Carbon receive?
It scored 59 points on the Artificial Analysis Intelligence Index v5, one point ahead of Claude Opus 5.5 at 58.
Does Carbon lead every benchmark in the index?
No. The source report says Carbon ranks first on four of nine tests and trails on five, including office-style knowledge work and real-world coding.
Can the public use Gemini 4 Carbon now?
Not according to the supplied report. Google is starting with a subset of cyber defenders; it has not provided a public release date.
What does the one-point lead say about Carbon?
It says Carbon scored highest under the index’s particular tests, weights and settings. It does not establish that Carbon is the best choice for every task or user.
How much is Carbon reported to cost?
The source lists introductory API prices of $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20. It does not give a date for the price change.
Source: ThorstenMeyerAI.com
