🔍 Read the full analysis: Gemini 4 Carbon’s One-Point AI Lead: What The Score Leaves Out on ThorstenMeyerAI.com
Get hardware and tech essentials 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 above Claude Opus 5.5 and two above GPT-6 Astra. It led on four of nine tests, but the index does not establish a universal best model: Carbon lost on five tests, carries different costs, and is not yet publicly available.
Gemini 4 Carbon has taken the top position on Artificial Analysis’ Intelligence Index v5 with 59 points, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The result gives Google’s model the highest aggregate score in this benchmark, but it is not a verdict on which AI is best for every task: Carbon placed first on four of nine tests, lost the other five, and is not yet available to the general public.
The index combines nine tests into one score, splitting its weight evenly across knowledge work, coding, science and general performance. Artificial Analysis launched version 5 on October 27. Its scores put Gemini 4 Carbon, in high-reasoning mode, at 59; Claude Opus 5.5, at its maximum setting, at 58; and GPT-6 Astra, also at maximum, at 57. Google’s previous model, Gemini 4 Argon, scored 50.
Carbon’s strongest results include HLE-Diamond, a hard reasoning and knowledge test, and Terminal-Bench-Science, which assesses science workflows conducted through a terminal. It also led AA-Omniscience, a knowledge test that penalizes wrong answers, and AutomationBench-AA, which tests business workflows across software tools. The source reports Carbon at 69% on HLE-Diamond, compared with 68% for Claude Opus 5.5, and at 65% on Terminal-Bench-Science, against 63% for GPT-6 Astra.
The overall ranking masks losses in other areas. Carbon trails Claude models on office-style knowledge work, Claude Opus 5.5 on real-world coding, GPT-6 Astra on physics research, Kimi K3.1 Pro on terminal tasks, and Meta’s Muse Wonder Full on document reasoning. The source lists a standard-price cost of $7.76 per index task: less than Claude Opus 5.5’s $10.04, but more than GPT-6 Astra’s $6.24 and GPT-6.1 Sol’s $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 Winner
A one-point gap matters because it places Carbon at the top of this particular aggregate ranking, but it is small relative to the range of tasks the score compresses. The index gives equal weight to four broad areas, while individual tasks within those areas carry different weights. A business choosing a model for coding, document analysis or office work may get a different answer from the headline ranking.
The result also highlights a practical distinction between accuracy and caution. On AA-Omniscience, Carbon scores 58% on plain accuracy, below several Claude and GPT models. It nevertheless leads the penalty-adjusted test because Artificial Analysis reports an 11% hallucination rate for Carbon, compared with about 50% for Claude Opus 5.5 and GPT-6 Astra. That result suggests fewer wrong answers among questions it misses, not that Carbon knows more or is reliable in every setting.
For companies, benchmark placement alone does not establish value. The listed task costs differ substantially, and the source’s cost figures do not describe every organization’s usage or workflow. Buyers will need to test their own workloads, including output quality, error rates, speed and total cost. For ordinary users, the immediate effect is limited: Carbon has not been opened to the public.
Products Worth Considering
Ledger Nano X Bluetooth Crypto Wallet
Apple MacBook Pro 2023 14 with M3 Pro Chip, 18 GB RAM, 2 TB SSD, Space Black
Apple MacBook Pro (14.2-inch, 2025) – M5 chip, 16GB Unified Memory, 1TB SSD, Space Black
As an affiliate, we earn on qualifying purchases.
How Artificial Analysis Builds the Score
Artificial Analysis describes its Intelligence Index as a combined score across nine benchmarks, rather than a direct measure of general intelligence. In version 5, knowledge work, coding, science and general tests each account for 25%. The source says the index uses long-running tasks such as agent work in terminals, report creation and code repair, rather than only short question-and-answer prompts.
According to the source, 60% of version 5 uses private test sets or held-out answers, intended to reduce the chance that models have memorized test questions. Even with those design choices, results apply to the benchmarks and model settings used. Artificial Analysis’ separate Coding Agent Index, which evaluates models alongside coding tools, is also reported to place Claude Code with Claude Opus 5.5 ahead of Google’s Antigravity CLI with Carbon. That comparison reinforces that a model’s performance can vary when the tool environment changes.
Google’s announced rollout begins with a subset of cyber defenders, then is planned to extend to trusted partners, developers, businesses and consumers. The source gives no dates for those later stages. It reports introductory API pricing of $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20 respectively.
““number one” on one index is a narrower claim than it sounds.”
— Thorsten Meyer, ThorstenMeyerAI.com
What the Ranking Cannot Settle
The source describes the benchmark results and rollout but does not provide the full underlying test data, independent replications or details of how the results might translate to particular customer workloads. It is not clear when Carbon will reach trusted partners, developers, businesses or consumers, or whether the reported API prices and rollout schedule will change.
Some comparisons also use different model settings, including high reasoning for Carbon and maximum settings for competitors. The source does not explain how that setting variation affects cost or performance across every test. Its account of Artificial Analysis’ comment on the coding-agent comparison is incomplete, ending mid-quotation, so the full statement cannot be established from the supplied material.
The figures should be read within their stated test conditions. A score of 59 does not establish that Carbon will outperform rivals on every real-world task, and the reported hallucination rates do not establish how often any model will make errors in a particular deployment.
Rollout and Real-World Testing
Google’s reported plan is to offer Carbon first to a limited group of cyber defenders, followed by trusted partners and then wider access for developers, businesses and consumers. No timetable for those stages is specified in the source. Further details on availability, final pricing and the conditions for broader access remain pending.
For organizations evaluating the model, the next useful evidence will come from testing Carbon on their own tasks and comparing finished-work quality, error handling, latency and total cost against alternatives. The index offers a starting point for those comparisons, not a substitute for them. Consumers will need to wait for Google to open access before they can assess whether the benchmark lead changes their experience.
Key Questions
What did Gemini 4 Carbon score?
It scored 59 on 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 the ranking mean Carbon is the best AI model?
No. It ranks models on nine specific tests combined into one score. Carbon led four tests and lost five, so another model may be a better fit for a particular task.
Why did Carbon lead the hallucination-penalty test?
Artificial Analysis reported that Carbon scored higher on AA-Omniscience because the test penalizes wrong answers and Carbon had a lower reported hallucination rate. Its plain accuracy score was below several Claude and GPT models, so the result reflects cautious answering rather than the highest number of correct answers.
Can the public use Gemini 4 Carbon now?
Not according to the supplied report. Google is starting with a subset of cyber defenders and plans later access for partners, developers, businesses and consumers, but no dates are given.
How much does Carbon cost to run?
The source reports $7.76 per index task at standard prices and API pricing of $2 per million input tokens and $10 per million output tokens initially, later rising to $4 and $20. Actual costs can vary by usage and task. AI services can produce errors, and organizations should account for those risks and costs when evaluating them.
Source: ThorstenMeyerAI.com
