🔍 Read the full analysis: How To Interpret Gemini 4 Carbon’s One-Point Lead As An AI Model 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. The benchmark source says Carbon leads four of nine tests, but the score is not a universal measure of model ability; Carbon is not yet publicly available and its rollout starts with selected cyber defenders.
Google’s Gemini 4 Carbon scored 59 points on the Artificial Analysis Intelligence Index v5, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra, according to the benchmark results reported by ThorstenMeyerAI.com on October 29. The narrow lead puts Carbon first on this particular nine-test index, but does not establish it as the best model for every task: it leads four tests, trails on five, and is not yet available to the public.
The index combines nine tests across knowledge work, coding, science and general ability, with each category carrying equal weight. Carbon’s score is listed at its high-reasoning setting; the results put Claude Opus 5.5 at its maximum setting at 58, GPT-6 Astra at 57 and GPT-6.1 Sol at 55. Google’s preceding model, Gemini 4 Argon, scored 50.
Carbon takes first place on HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. It falls behind on office-style knowledge work and real-world coding, as well as the reported physics-research, terminal-task and document-reasoning comparisons. The source says the overall score reflects performance across this specific test mix, not a single measure of general intelligence.
The reported cost is $7.76 per index task at standard prices. That is below Claude Opus 5.5’s reported $10.04, but above GPT-6 Astra’s $6.24 and far above GPT-6.1 Sol’s $1.27. The source also reports an introductory API price of $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20, respectively.
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 a model, the result is a reason to examine Carbon, not to switch based on rank alone. A one-point margin on a combined benchmark can conceal much larger differences on particular tasks: Carbon performs well on the reported science and abstention-scored knowledge tests, while Claude models lead the office-work and coding comparisons. The useful question is whether those strengths match a buyer’s actual workload.
Cost also changes the comparison. A price per benchmark task does not by itself predict a company’s bill, which depends on task length, token use, retries and the tools surrounding the model. Still, the reported figures show why buyers should test cost per successfully completed task, alongside accuracy and speed, rather than treat the headline rank as a purchasing recommendation.
For individual users, there is no immediate product change because Carbon has not been released publicly. If it reaches general use, the source expects its advantages to matter more on long, difficult tasks than routine chat. That is an interpretation, not a confirmed user-experience result; independent access and testing would be needed to establish how it performs in everyday use.
Products Worth Considering
DJI Osmo Pocket 3 Creator Combo Vlogging Camera with Mic 2, 1 CMOS, 4K/120fps, 3-Axis Stabilization, 128GB Card
Jackery HomePower 3000 Portable Power Station with Two 200W Solar Panels
Apple iPad Air (4th Generation, 10.9-inch, Wi-Fi, 256GB) – Space Gray
As an affiliate, we earn on qualifying purchases.
How the Index Builds Its Score
Artificial Analysis launched Intelligence Index v5 on October 27, according to the source article. Its nine tests are divided into four equally weighted blocks: knowledge work, coding, science and general performance. Within each block, individual tests carry different shares of the total score, so a model’s result depends partly on which tasks the index includes and how heavily each is weighted.
The source reports that 60% of v5 uses private test sets or held-out answers, and that many tasks require models to act over longer workflows, such as using a terminal, preparing reports or modifying code. This design aims to test more than short question-and-answer performance. It also means benchmark results may reflect an agent’s ability to complete a workflow, not just the underlying model in isolation.
One result illustrates how scoring rules affect rankings. Carbon reportedly gets 58% of questions right on AA-Omniscience, below several competing models on raw accuracy, but its score rises because the test penalizes wrong answers and Carbon is reported to abstain more often than those rivals. The source gives Carbon an 11% hallucination rate under that test’s definition, compared with about 50% for Claude Opus 5.5 and GPT-6 Astra. Those figures describe this benchmark and should not be treated as universal error rates.
“Carbon wins four of the nine tests behind the score and loses the other five.”
— Thorsten Meyer, ThorstenMeyerAI.com
Access, Speed and Real-World Fit
Google has not provided a public release date in the details supplied by the source. The described sequence begins with selected cyber defenders, then trusted partners, followed by developers, businesses and consumers, but the schedule and eligibility criteria are unspecified.
The source says Carbon is slow to respond, but the provided material gives no latency measurement or direct comparison figures. It also does not supply enough detail to independently assess the benchmark runs, including the full methodology for every test, uncertainty ranges around the one-point margin, or how results vary across repeated runs. The score should be read as a reported ranking, not as proof that the models differ reliably in every setting.
Some comparisons mentioned in the source are incomplete in the material provided, including the account of the Coding Agent Index. The headline index also tests a defined selection of tasks; it cannot establish how Carbon will perform across every company’s software, data, security requirements or user-facing applications.
Google’s Staged Access Rollout
The next stated step is access for a subset of cyber defenders, followed by trusted partners and later broader groups, according to the source. Google has not given dates for those stages. Wider access would allow developers and organizations to test Carbon directly against their own workloads and compare quality, response time and total operating costs.
For now, the practical next step for buyers is to wait for access and evaluate representative tasks rather than rely on a single overall score. Further benchmark results, clear release timing and independent reports from users would help show whether Carbon’s one-point lead translates into an advantage outside the index.
Key Questions
What does Gemini 4 Carbon’s 59-point score mean?
It is Carbon’s reported score on Artificial Analysis Intelligence Index v5, a combined ranking based on nine tests across four equally weighted categories. It does not measure every ability or use case.
How far ahead is Carbon?
Carbon’s reported score of 59 is one point above Claude Opus 5.5 at 58 and two above GPT-6 Astra at 57. The source does not provide uncertainty ranges for the difference.
Does Carbon lead every benchmark test?
No. The source says Carbon ranks first on four of the nine tests and trails on five, including comparisons involving office-style work and real-world coding.
Can the public use Gemini 4 Carbon now?
Not according to the source article. Google is described as starting with selected cyber defenders and expanding access in later stages, but no public release date is given.
Should businesses switch to Carbon because it ranked first?
The ranking alone is not enough to justify a switch. Organizations should test their own workloads and compare accuracy, completion time and cost per finished task, since Carbon’s relative performance varies across benchmark categories.
Source: ThorstenMeyerAI.com
