Read the full analysis: How Gemini 4 Carbon Reached The Top Of The AI Ranking By One Point on ThorstenMeyerAI.com
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 four of the index’s nine tests, but trails competitors on five; Google says access will begin with a limited group of cyber defenders, and has not given a public release date.
Google’s Gemini 4 Carbon has taken first place in the Artificial Analysis Intelligence Index v5 with a score of 59, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. The result puts Carbon at the top of this particular nine-test benchmark, but the model is not yet publicly available: Google is beginning access with a small group of cyber defenders, with no date announced for wider release.
Artificial Analysis rebuilt its leaderboard on October 27, and Claude Opus 5.5 initially ranked first. Carbon took the lead the next day, according to the source report. In the listed scores, Claude Opus 5.5 (maximum reasoning) has 58 points, GPT-6 Astra (maximum) 57, GPT-6.1 Sol (maximum) 55, and Google’s earlier Gemini 4 Argon 50. Carbon’s reported 59-point score is for its high-reasoning setting.
Carbon placed first on four of nine tests: HLE-Diamond, a difficult reasoning and knowledge test; Terminal-Bench-Science, which evaluates science workflows in a terminal; AA-Omniscience, which penalizes incorrect answers; and AutomationBench-AA, a business-workflow test. It did not lead the other five. Claude models topped office-style knowledge work, Claude Opus 5.5 led the real-world 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 report.
The report puts Carbon’s cost at $7.76 per index task at standard prices. That is below the stated $10.04 for Claude Opus 5.5 but above GPT-6 Astra’s $6.24 and GPT-6.1 Sol’s $1.27. Google’s introductory API pricing is listed as $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20 respectively. Those figures do not, by themselves, establish what a typical customer will pay for a finished task.
Index v5
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.
Hard reasoning + knowledge questionsWeight 10%Carbon 69%Carbon leads Next: Claude Opus 5.5, 68%Terminal-Bench-Science
Science research work in a terminalWeight 15%Carbon 65%Carbon leads Next: GPT-6 Astra, 63%AA-Omniscience v1.1
Knowledge, wrong answers penalizedWeight 15%Carbon 54Carbon leads Next: Claude Opus 5.5, 48AutomationBench-AA
Business workflows across appsWeight 5%Carbon 78%Carbon leads Next: Gemini 4 Argon, 78%AA-Briefcase
Office-style knowledge workWeight 15%Carbon 1,609 EloBehind Claude Sonnet 5.5, 1,812FrontierCode 2.0
Code a maintainer would mergeWeight 15%Carbon 41%Behind Claude Opus 5.5, 46%CritPt v2 (hard)
Research-level physicsWeight 10%Carbon 63%Behind GPT-6 Astra, 71%Terminal-Bench 4.0
Agentic terminal tasksWeight 10%Carbon 65%Behind Kimi K3.1 Pro, 67%GDP.pdf
Reasoning over documentsWeight 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, generalCarbon trails in two of the four heaviest tests (15% each): office-style work and real-world codeHallucination 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 Verdict
The result matters because leaderboard positions can shape how companies compare models and decide which systems to test. But the one-point margin is a small difference on a composite score, not evidence that Carbon is best for every use. Its uneven results suggest buyers should match a model to specific work rather than choose solely by overall rank.
Cost and reliability also affect the practical choice. Carbon’s price per benchmark task is lower than Claude Opus 5.5’s in the comparison, but higher than the two listed GPT alternatives. On AA-Omniscience, Carbon’s reported advantage comes from avoiding wrong answers rather than having the highest raw accuracy: its accuracy is 58%, compared with 66% for Claude Opus 5.5 and 62% for GPT-6 Astra. The report gives Carbon an 11% hallucination rate on that test, against about 50% for those two models. That may make its cautious answering useful in settings where unsupported answers are costly, though the benchmark does not prove how it will perform in every real-world setting.
For everyday users, there is no immediate product change while public access remains unavailable. For businesses, the ranking is best treated as a reason to run task-specific evaluations, including the cost of completing work, accuracy, and how often a system declines to answer. The score alone does not settle those questions.
How the Nine-Test Index Works
Artificial Analysis describes its v5 index as a combined score across nine tests divided into four equally weighted blocks: knowledge work, coding, science, and general capability. The source report says 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 CritPt v2; and general capability includes AA-Omniscience and HLE-Diamond. Individual tests carry different weights within the blocks.
The report says 60% of the index uses private test sets or held-out answers, intended to reduce the chance that models can rely on memorized questions. Many tasks also involve extended work, such as using terminal tools, building reports, or changing code. That gives the index a different emphasis from a simple question-and-answer exam, but it also means the final order reflects the specific tasks, settings, and scoring choices in v5.
One example is AA-Omniscience. Artificial Analysis’s measure rewards correct answers and penalizes wrong ones. The report says Carbon has lower raw accuracy than several rivals but scores better on that test because it makes fewer incorrect guesses. Its result therefore reflects a balance between answering correctly and abstaining, not a claim that it possesses the broadest knowledge.
Release Dates and Real-World Results
No public launch date is specified for Gemini 4 Carbon. The source report says access will expand from cyber defenders to trusted partners, developers, businesses, and consumers, but does not give a schedule or explain how users will qualify for each stage.
The available material also does not show how Carbon performs across the full range of customer workloads, or whether the benchmark’s one-point lead is statistically meaningful. The ranking compares selected models at stated reasoning settings on a particular set of tests; results could differ with other tasks, tools, settings, or pricing arrangements. The report’s comparison of cost per index task should not be read as a universal estimate of operating costs.
Some benchmark details and figures are presented through the source report rather than full underlying test data in the supplied material. Readers should treat the score as a reported result from Artificial Analysis, not as an independent conclusion that Carbon is broadly superior to rival systems.
Google’s Staged Access Plan
Google’s stated plan is to begin with a limited group of cyber defenders, then extend access to trusted partners, developers, businesses, and consumers. The next concrete development would be more information on eligibility, timing, and availability through Google’s products or API; none is specified in the source report.
In the meantime, the ranking can inform model evaluations, but organizations considering a switch will need to test Carbon against their own workloads and compare quality, completion time, reliability, and cost. For individual users, the immediate status is unchanged: Carbon is not yet available for general use.
Key Questions
What score did Gemini 4 Carbon receive?
It received 59 points on Artificial Analysis’s Intelligence Index v5, according to the source report. Claude Opus 5.5 scored 58 and GPT-6 Astra scored 57.
Does the ranking mean Carbon is the best AI model for every task?
No. The score combines results from nine specific tests. Carbon led four, while other models led the remaining five, including tests for office-style work and coding.
Can the public use Gemini 4 Carbon now?
No public availability is reported. Google is starting with a limited group of cyber defenders and has not announced dates for broader access.
How much does Carbon cost?
The report lists $7.76 per index task at standard prices and introductory API rates of $2 per million input tokens and $10 per million output tokens. It says those rates later rise to $4 and $20, respectively.
Why did Carbon lead the AA-Omniscience test?
The test penalizes wrong answers. The report says Carbon’s 11% hallucination rate helped it score well despite lower raw accuracy than several competitors, because it was less likely to give an incorrect answer.
Source: ThorstenMeyerAI.com