🔍 Read the full analysis: Gemini 4 Carbon’s AI Ranking Win Is Narrow; Here’s What The Score Omits 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 leads Artificial Analysis’ Intelligence Index v5 with 59 points, one ahead of Claude Opus 5.5. The ranking combines nine tests: Carbon leads four and trails rivals on five, while its cost and limited rollout complicate what the top position means.
Google’s Gemini 4 Carbon now leads the Artificial Analysis Intelligence Index v5 with 59 points, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra, according to figures reported by ThorstenMeyerAI.com. The result puts Carbon first on this particular nine-test benchmark, but it does not make the model the top performer on every task: the report says it wins four tests, loses five, and is not yet broadly available.
The index, rebuilt by the independent benchmarking firm Artificial Analysis and launched on October 27, combines nine tests into a single score. The report lists Claude Opus 5.5 at 58 points, GPT-6 Astra at 57, GPT-6.1 Sol at 55 and Google’s previous model, Gemini 4 Argon, at 50. Carbon’s reported result uses its high-reasoning setting; the rival scores cited use the settings specified in the source.
Carbon takes first place in four tests: HLE-Diamond, a difficult knowledge and reasoning exam; Terminal-Bench-Science, which tests science workflows in a terminal; AA-Omniscience, which penalizes incorrect answers; and AutomationBench-AA, focused on business workflows across software tools. Its reported results include 69% on HLE-Diamond, compared with Claude Opus 5.5’s 68%, and 65% on Terminal-Bench-Science, compared with GPT-6 Astra’s 63%.
The model trails competitors on the other five tests, according to the source report. Claude models lead on office-style knowledge work, Claude Opus 5.5 leads on real-world coding, GPT-6 Astra leads on physics research, Kimi K3.1 Pro leads on terminal tasks, and Meta’s Muse Wonder Full leads on document reasoning. The report also puts Carbon’s estimated cost at $7.76 per index task at standard prices. That is below the cited $10.04 for Claude Opus 5.5, but above 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 One-Point Lead, Not a Universal Win
The ranking offers a useful comparison of models on a defined mix of long, demanding tasks, but its one-point margin is narrow and its overall score combines areas where Carbon has different strengths. For organizations choosing a model, the test-level results and the cost of completing their own work may be more relevant than the overall order. The reported price per index task is a benchmark estimate, not a guarantee of what a specific company will pay in production.
The source says Carbon’s strongest distinction on AA-Omniscience is not raw answer accuracy but caution. It reports Carbon at 58% accuracy on that test, below several Claude and GPT models, while recording an 11% hallucination rate. The test rewards avoiding wrong answers, so a model that declines to guess can score better even if it answers fewer questions correctly. That is a meaningful difference for tasks where unsupported answers carry costs, but it does not establish that Carbon is more accurate across other kinds of work.
For consumers, the immediate effect is limited: the report says the model is not yet open to general users. For businesses and developers, the ranking is an early signal rather than a reason on its own to switch systems. Performance, reliability, availability and costs on an organization’s specific workload remain key considerations.
Products Worth Considering
DJI Mic 3 Wireless Microphone System (2 Transmitters, 1 Receiver, Charging Case)
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
EcoFlow Smart Home Panel 2 Whole-Home Backup System Hub
As an affiliate, we earn on qualifying purchases.
How Artificial Analysis Builds the Score
Artificial Analysis’ version 5 index divides its tests into four equally weighted blocks: knowledge work, coding, science and general ability, each worth 25% of the combined score. The report assigns the largest individual weights to AA-Briefcase, FrontierCode 2.0, Terminal-Bench-Science and AA-Omniscience, each at 15%. Other tests include AutomationBench-AA, GDP.pdf, Terminal-Bench 4.0, CritPt v2 and HLE-Diamond.
The source says 60% of the index uses private test sets or held-out answers, intended to reduce the effect of models having seen test questions during training. Many evaluations also involve extended work, such as using a terminal, producing reports or modifying code. Those methods make the benchmark different from a short question-and-answer test, while also making runs more costly.
The source describes the index as a measure of how a model performs at a particular setting on these nine tests, not a definitive ranking of intelligence. Its report also cites introductory API pricing for Carbon of $2 per million input tokens and $10 per million output tokens, with later prices listed as $4 and $20, respectively. Those prices and the per-task cost estimate describe different measures and should not be treated as interchangeable.
““Carbon wins four of the nine tests behind the score and loses the other five.””
— ThorstenMeyerAI.com report
Availability and Benchmark Limits
Google has not given a public release date in the source material. It says Carbon is being offered first to “a subset of cyber defenders”, followed by trusted partners and then developers, businesses and consumers. The size of the initial group, the timing of later access and whether all users will receive the same model settings are not stated.
The ranking figures here come from the supplied ThorstenMeyerAI.com report; the underlying test data, run-to-run variation and full methodology are not independently reproduced in that material. The provided source text also cuts off during its discussion of the coding comparison, so further explanation from Artificial Analysis or Google is not available here. It remains unclear how Carbon’s performance will compare on individual customer workloads, how often results might change with model or tool updates, and what final pricing will apply to each access route.
Google’s Staged Rollout Ahead
According to the report, Google plans to move Carbon from a limited group of cyber defenders to trusted partners, then to developers, businesses and consumers. No schedule is specified, so access timing is still uncertain. The rollout and any subsequent public pricing details will determine when users can assess the model directly rather than through benchmark results.
For organizations evaluating Carbon, the next useful evidence will be performance on their own tasks, including the quality of completed work, the frequency of incorrect answers, response time and cost. Future benchmark updates may also show whether the narrow lead persists as models and tests change. Until then, the reported 59-point result is best understood as a snapshot of one model configuration on one index.
Key Questions
What is Gemini 4 Carbon’s score?
The source report says Carbon scored 59 points on Artificial Analysis’ Intelligence Index v5, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra.
Does Carbon lead on every benchmark test?
No. The report says Carbon ranks first on four of nine tests and trails other models on five, including office-style knowledge work and real-world coding.
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 to expand access in stages, but has not announced a public release date.
Why does Carbon score well on the hallucination test?
The report says AA-Omniscience penalizes wrong answers. Carbon has lower reported raw accuracy than several rivals on that test, but its reported hallucination rate is also lower, which can help on a scoring system that rewards caution.
How much does Carbon cost to run?
The report estimates $7.76 per index task at standard prices and lists introductory API prices of $2 per million input tokens and $10 per million output tokens. Actual costs depend on usage and the pricing that applies when access is available.
Source: ThorstenMeyerAI.com
