🔍 Read the full analysis: Does A One-Point Lead Make Gemini 4 Carbon The Best AI Model For You? on ThorstenMeyerAI.com
Get smart everyday buys 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. It leads on four of nine tests but trails rivals on others, costs more than some competitors per index task and is not yet publicly available.
Google’s Gemini 4 Carbon has taken the top spot 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 not a clear all-purpose winner: it places first on four of nine tests, trails competitors on several others and remains unavailable to the public.
Artificial Analysis, an independent benchmarking company, released version 5 of its index on October 27. The reported scores put Claude Opus 5.5 at 58, GPT-6 Astra at 57, GPT-6.1 Sol at 55 and Google’s earlier Gemini 4 Argon at 50. Carbon’s lead is therefore narrow and specific to this index, rather than a decisive margin across every kind of work.
Carbon ranks first on tests for hard reasoning and knowledge questions, science workflows performed in a terminal, knowledge questions that penalize wrong answers, and business workflows across software tools. It scores 69% on HLE-Diamond, one percentage point above Claude Opus 5.5, and 65% on Terminal-Bench-Science, two points above GPT-6 Astra. On the business-workflow test, AutomationBench-AA, it reaches 78%, slightly ahead of Gemini 4 Argon.
The model’s reported price is $7.76 per index task at standard prices. That is below Claude Opus 5.5’s $10.04 but above GPT-6 Astra’s $6.24 and substantially above GPT-6.1 Sol’s $1.27. The source also lists an introductory API price of $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20, respectively. The per-task and token prices describe different measures and should not be treated as interchangeable.
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 a One-Point Lead Can Tell Buyers
The ranking may help companies identify models to test, but its narrow margin does not settle which system is best for a particular job. The index combines results across knowledge work, coding, science and general tasks, with each category given equal weight. A buyer whose work is concentrated in coding or office-style document tasks may care more about those individual results than the overall score.
Cost also changes the practical comparison. Carbon is cheaper per index task than Claude Opus 5.5 in the figures provided, but more expensive than GPT-6 Astra and far more expensive than GPT-6.1 Sol. Those amounts are benchmark-run costs, not a guarantee of what any company’s own workload will cost. Businesses would need to test models on their own tasks and compare the price of completed work, including the quality and amount of human review required.
For individual users, the ranking does not signal an immediate product change: Carbon is not yet publicly available. If and when it reaches general users, its benchmark strengths are more relevant to lengthy reasoning, research and tool-based tasks than to a claim that it is best at every everyday chat.
Products Worth Considering
Dell UltraSharp 32 U3226QER 4K Thunderbolt Hub Monitor
FLEXISPOT 42-Inch Electric Height-Adjustable Desk Converter with Keyboard Tray
FLEXISPOT Standing Desk with Monitor Riser & Cable Management
As an affiliate, we earn on qualifying purchases.
How Artificial Analysis Builds the Score
Version 5 combines nine benchmarks into a single score, grouping them into four equally weighted categories. 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. The weights differ within each category, so a result on a heavily weighted test can affect the total more than one on a smaller test.
According to the source account, 60% of the index uses private test sets or held-out answers, and many tasks require models to act over extended workflows, such as operating in a terminal, preparing reports or modifying code. That design aims to test more than short answers, but the index remains a result for these nine tasks, at particular model settings. The source says the leaderboard was rebuilt the day before Google’s model took the lead; it does not provide enough detail here to independently assess the full change in methodology or every model’s testing conditions.
Carbon’s result on AA-Omniscience illustrates how benchmark design affects a ranking. It scored 54 on the test’s measure, six points ahead of Claude Opus 5.5, while its plain answer accuracy was 58%. The source reports higher accuracy for Claude Fable 5.1 at 68%, Claude Opus 5.5 at 66%, and GPT-6.1 Sol and GPT-6 Astra at 62% each. Carbon’s advantage on the penalized test comes from its reported tendency to give fewer incorrect answers when it does not know: its hallucination rate was 11%, compared with about 50% for Claude Opus 5.5 and GPT-6 Astra. The result rewards restraint, not the highest number of correct answers.
Carbon is not first across all tested work. The source reports that Claude Sonnet 5.5 and Claude Opus 5.5 lead on AA-Briefcase, an office-style knowledge-work test, while Claude Opus 5.5 leads on real-world coding. GPT-6 Astra leads the cited physics research test, Kimi K3.1 Pro the terminal-task test, and Meta’s Muse Wonder Full the document-reasoning test. These task-level differences are why the overall order should be read alongside the underlying results.
What the Ranking Does Not Resolve
The one-point difference between Carbon and Claude Opus 5.5 is the reported index result; the supplied material does not give uncertainty ranges or repeated-run variation. It is therefore unclear how stable that small gap would be under different runs or testing conditions. The index also cannot establish a universal ranking: changing the task mix, weights or model settings could change the order.
Some comparisons in the source are summaries rather than full benchmark tables. The material does not provide complete results for every model on every test, nor details sufficient to compare all deployment settings. It also does not state a date for broad availability. Google is reported to be offering Carbon first to a subset of cyber defenders, then to trusted partners, developers, businesses and consumers, but the schedule and eligibility at each stage remain unspecified.
The benchmark’s cost figure does not tell readers what Carbon will cost for a particular production workload. Actual spending depends on task length, use of input and output tokens, model settings and how often a system must retry or hand work to a person. The quoted API prices are reported figures, and the source says they are due to change; it does not provide a date for that change.
When Carbon May Reach More Users
Google’s reported rollout begins with a limited group of cyber defenders, followed by trusted partners, then developers, businesses and consumers. No public release date is given in the source material, so it is not possible to say when ordinary users will be able to try Carbon or when each stage will begin.
For organizations considering the model, the next practical step is to run trials against their own work rather than select a system solely from the leaderboard. Those tests can compare accuracy, refusals, speed, tool use and cost per completed task. Until broader access and real-world results are available, the index offers a benchmark snapshot—not a complete guide to how Carbon will perform in a company’s workflow or a consumer product.
Key Questions
Is Gemini 4 Carbon the best AI model overall?
It leads the Artificial Analysis Intelligence Index v5 with 59 points, but that result applies to the index’s nine tests and weighting. Carbon does not lead every benchmark, so the score does not establish that it is best for every use.
How far ahead is Carbon?
Carbon scores one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra in the reported index results. The supplied material does not give uncertainty ranges for the scores.
Why does Carbon win the knowledge test if other models answer more questions correctly?
AA-Omniscience penalizes wrong answers. The source reports that Carbon has lower plain accuracy than several competitors but a lower reported hallucination rate, so it can score better on a test that rewards not guessing when uncertain.
Can the public use Gemini 4 Carbon now?
No public access is reported yet. Google is said to be starting with a subset of cyber defenders and expanding to other groups, but no timeline is provided.
How much does Carbon cost?
The source reports a standard benchmark cost of $7.76 per index task, plus introductory API prices of $2 per million input tokens and $10 per million output tokens. It says those API prices will later rise to $4 and $20, respectively, but gives no date. A company’s actual costs may differ by workload.
Source: ThorstenMeyerAI.com
