🔍 Read the full analysis: Gemini 4 Carbon’s Narrow AI Model Lead: Five Things The Ranking Leaves Out 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’s Intelligence Index v5, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The ranking combines nine tests; Carbon leads four, trails on five, and is not yet publicly available. Its price and task-specific performance matter more than the overall position for users deciding whether to adopt it.
Google’s Gemini 4 Carbon scored 59 points to lead the Artificial Analysis Intelligence Index v5, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The result puts Carbon atop this particular nine-test ranking, but it does not mean the model leads on every task: it won four tests, lost five, and is not yet available to the public.
The index, rebuilt by independent benchmarking company Artificial Analysis and launched October 27, gives equal weight to four categories: knowledge work, coding, science and general performance. Carbon ranked first on HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. It scored 69% on HLE-Diamond, against 68% for Claude Opus 5.5, and 65% on Terminal-Bench-Science, against 63% for GPT-6 Astra.
Carbon’s results were weaker in several other areas. Claude Sonnet 5.5 and Claude Opus 5.5 scored higher on office-style knowledge work, while Opus 5.5 led on the FrontierCode 2.0 coding test. GPT-6 Astra led the physics-research test, Kimi K3.1 Pro led terminal tasks, and Meta’s Muse Wonder Full led document reasoning, according to the source’s account of the benchmark results. The overall score therefore compresses sharply different task results into a single ranking.
The reported standard-price cost is $7.76 per index task. That is below the cited $10.04 for Claude Opus 5.5 at its maximum setting, but above GPT-6 Astra’s $6.24 and GPT-6.1 Sol’s $1.27. Google lists introductory API pricing at $2 per million input tokens and $10 per million output tokens, with later pricing of $4 and $20 respectively, according to the source. The cost of a benchmark task is not the same as a company’s cost per completed job.
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
What the One-Point Lead Shows
The result matters because companies often use model rankings to shortlist systems for work such as coding, research and handling documents. But a one-point overall margin does not establish that Carbon is the best choice for a particular workload. The index’s four equally weighted categories and nine selected tests determine the outcome; different tasks, settings or cost limits could change which model is most useful.
Carbon’s AA-Omniscience result also illustrates how a benchmark’s scoring rules shape what “leading” means. The source reports that Carbon answered 58% of questions correctly, below several Claude and GPT models, but scored best because the test penalizes incorrect answers and Carbon more often refrained from guessing. Its reported hallucination rate was 11%, compared with about 50% for Claude Opus 5.5 and GPT-6 Astra. That is a measure of performance under this test’s definition, not proof that Carbon will be more reliable in every real-world use.
Availability is another practical constraint. Google is first offering Carbon to a subset of cyber defenders, then plans staged access for trusted partners, developers, businesses and consumers. No schedule was provided in the source. For now, the ranking is relevant to buyers evaluating future options, not a signal that ordinary users can switch to Carbon today.
Products Worth Considering
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Jade
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
Apple iPad Air 11-inch (M2), 512GB, Wi-Fi, Space Gray
As an affiliate, we earn on qualifying purchases.
How Artificial Analysis Built the Index
Artificial Analysis describes v5 as a combined score across nine tests, with each of its four blocks carrying 25% of the total. 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 general performance includes AA-Omniscience and HLE-Diamond.
According to the source, 60% of the tests use private test sets or held-out answers, an effort to reduce the effect of models having encountered questions during training. Many tasks also involve longer agent workflows, such as working in a terminal, producing reports or changing code. Those design choices aim to test more than short question-and-answer performance, though they also make results harder to compare with unrelated benchmarks.
The reported scores place Gemini 4 Argon, Google’s previous model, at 50 points, compared with Carbon’s 59. Claude Opus 5.5 scored 58, GPT-6 Astra 57, and GPT-6.1 Sol 55. These figures apply to the specific model versions and settings named in the source, rather than to every version or configuration from those companies.
Limits of the Ranking and Rollout
The source describes Carbon as first in the index but supplies no full table for all nine tests in the material provided. Its account says Carbon wins four and loses five, while naming several losses; the complete test-by-test scores and the size of every margin are not included. Readers should treat those specific comparisons as reported results, not as a full independent reconstruction of the leaderboard.
It is also unclear when Carbon will reach developers, businesses or consumers. Google has not supplied a public release date in the source. The benchmark results do not establish how Carbon performs across a company’s own workflows, how its costs would compare on those workloads, or whether the reported hallucination-rate difference will hold outside AA-Omniscience.
The source also cites a separate Coding Agent Index comparing Claude Code with Claude Opus 5.5 against Google’s Antigravity CLI with Carbon, with scores of 62 and 59 respectively. That comparison includes tools as well as models, so it should not be read as a model-only result. The provided account ends before fully explaining Artificial Analysis’s assessment of that comparison.
Access and Real-World Testing
Google’s stated plan is to offer Carbon first to a subset of cyber defenders, then to trusted partners, followed by developers, businesses and consumers. The timing and size of each stage remain unspecified in the source. Any broader availability announcement would clarify when users can test the model directly and which products or API access routes will support it.
For prospective business users, the next useful evidence will come from tests on their own tasks: accuracy, completion time, need for human correction and cost per finished job. The index offers a way to compare performance across its chosen tests, but the one-point lead alone does not resolve those operational questions. Public users have no immediate change to make while Carbon remains unavailable to them.
Key Questions
What is Gemini 4 Carbon’s score?
It scored 59 points on Artificial Analysis’s Intelligence Index v5, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra, according to the source.
Does Carbon lead on every benchmark test?
No. The source says Carbon leads four of nine tests and loses the other five, including tests in office-style knowledge work and real-world coding.
Why can Carbon lead the omniscience test without answering the most questions correctly?
AA-Omniscience penalizes wrong answers. The source reports that Carbon’s lower accuracy than several rivals was offset by a lower reported hallucination rate, so it scored higher under that test’s scoring rules.
Can the public use Gemini 4 Carbon now?
Not according to the source. Google is starting with a subset of cyber defenders and has not given a public date for broader access.
Is Carbon cheaper than its competitors?
It depends on the comparison. The source gives a standard index-task cost of $7.76, below Claude Opus 5.5’s cited $10.04 but above GPT-6 Astra’s $6.24 and GPT-6.1 Sol’s $1.27. Real costs will depend on the workload and usage.
Source: ThorstenMeyerAI.com
