🔍 Read the full analysis: Before You Call Gemini 4 Carbon The Top AI Model, Check The Ranking Details 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 ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra. It leads four of nine tests, but the benchmark does not establish an overall best model for every task, and Google has not made Carbon publicly available.
Gemini 4 Carbon has taken first place 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 leader on this particular nine-test benchmark, not a universal winner: it places first on four tests, trails rivals on five, and is not yet publicly available.
Artificial Analysis, an independent benchmarking company, introduced version 5 on October 27. Its index gives equal weight to four areas—knowledge work, coding, science and general capability—and combines results from nine tests. The reported scores put Claude Opus 5.5 at 58, GPT-6 Astra at 57, GPT-6.1 Sol at 55 and Google’s previous model, Gemini 4 Argon, at 50.
Carbon leads on four of the nine tests: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. Its reported scores include 69% on HLE-Diamond, against 68% for Claude Opus 5.5, and 65% on Terminal-Bench-Science, against 63% for GPT-6 Astra. On AutomationBench-AA, which evaluates business workflows using software tools, Carbon narrowly outscored its predecessor.
It falls behind competitors on the other five tests, including office-style knowledge work, real-world coding, physics research, terminal tasks and document reasoning. On the index’s office-work measure, Carbon scored 1,609 Elo, compared with 1,812 for Claude Sonnet 5.5 and 1,800 for Claude Opus 5.5. On FrontierCode 2.0, it scored 41%, while Claude Opus 5.5 scored 46%.
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
Why the One-Point Lead Has Limits
The ranking can help organizations compare models, but the 59-point total compresses different strengths into one number. A one-point margin over Claude Opus 5.5 is small, and the result depends on the index’s selected tasks, their weights and the model settings tested. It should not be read as proof that Carbon performs best across every product, workflow or user’s needs.
Cost and access also affect the practical value of the result. The source reports a standard-price estimate of $7.76 per index task for Carbon—less than Claude Opus 5.5’s $10.04, but above GPT-6 Astra’s $6.24 and well above GPT-6.1 Sol’s $1.27. These figures describe the benchmark’s task costs; they do not show what a particular organization would pay for its own workload.
For people choosing a system, the useful comparison is performance on their actual tasks alongside cost, speed and reliability. A model that leads an aggregate benchmark may not be the strongest or least expensive option for a company’s coding, document or office work. The ranking is a comparison point, not a recommendation to adopt a particular model.
Products Worth Considering
Apple MacBook Pro 16.2-inch with M5 Pro, 18-core CPU, 20-core GPU, 48GB Memory, 1TB SSD, Wi-Fi 7, Space Black
Apple MacBook Pro 16-inch (2021) – M1 Pro Chip, 10-Core CPU, 16-Core GPU, 16GB RAM, 1TB SSD, Space Gray
LG 24G414B-B 24-Inch UltraGear Full HD IPS Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How Artificial Analysis Builds Its Index
Artificial Analysis describes the Intelligence Index as a combined score from nine tests in four equally weighted blocks. Knowledge work accounts for 25% of the total, as do coding, science and general capability. Within those blocks, individual tests have different weights: for example, AA-Briefcase contributes 15% and AutomationBench-AA 5%, while FrontierCode 2.0 contributes 15% and Terminal-Bench 4.0 contributes 10%.
The testing is intended to cover more than short question-and-answer exchanges. The source says many tasks involve extended work, such as operating in a terminal, creating reports or fixing code, and that 60% of v5 uses private test sets or held-out answers. Those choices may reduce the influence of memorized questions, but the index still measures performance on its own defined set of tasks—not every capability that matters in daily use.
One result needs particular care: Carbon leads AA-Omniscience, but the source says it did not have the highest raw accuracy. It reports Carbon at 58% accuracy and an 11% hallucination rate on that test, while several Claude and GPT models answered more questions correctly. The test also penalizes wrong answers, so a model that abstains rather than guessing can score better. The ranking therefore reflects both answer knowledge and willingness to risk an incorrect response.
““choose by task and by cost per finished task, not by rank.””
— Thorsten Meyer, source article
What the Benchmark Cannot Settle
The available information does not establish how Carbon will perform across the full range of real customer workloads or whether the small lead will persist as models and benchmark results change. The index compares specific models at specific settings on nine tests; different tasks, weights or model configurations could produce a different order. The source also reports scores and prices but does not provide enough detail here to independently reproduce every result or calculate costs for other usage patterns.
Public access timing remains unknown. Google is reported to be starting with selected cyber defenders, then planning access for trusted partners, developers, businesses and consumers, but no dates are given. The source mentions a planned introductory API price of $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20 respectively; the timing and conditions for those prices are not specified.
Google’s Staged Carbon Rollout
Google’s stated sequence is to provide Carbon first to a subset of cyber defenders, then to trusted partners, followed by developers, businesses and consumers. The company has not announced a date for those later stages in the source material. Until access expands, most users cannot test whether Carbon’s benchmark strengths translate to their own work.
When access becomes available, the relevant next comparison will be task-specific: results on an organization’s own coding, research, document or business processes, alongside the cost and time required to complete them. Further releases from Google and subsequent benchmark updates may clarify Carbon’s availability and whether its narrow lead holds across new evaluations.
Key Questions
Is Gemini 4 Carbon the best AI model overall?
The reported result establishes that Carbon leads Artificial Analysis’ Intelligence Index v5 with 59 points. It does not establish that Carbon is best for every task or evaluation; it wins four of the nine tests and trails other models on five.
How far ahead is Carbon?
Carbon scored 59 points, one point ahead of Claude Opus 5.5 at 58 and two ahead of GPT-6 Astra at 57. The ranking is for this index and its tested settings.
Can the public use Gemini 4 Carbon now?
Not according to the source. Google is reported to be starting with a subset of cyber defenders, followed by other groups in stages. No date for public availability is provided.
Why does Carbon lead the hallucination-penalized test?
The source reports that Carbon scores well on AA-Omniscience because the test penalizes wrong answers and Carbon is less likely to guess. It reports 58% raw accuracy and an 11% hallucination rate; several competing models have higher raw accuracy, so the test’s scoring method matters.
What does the index say about Carbon’s cost?
The source estimates $7.76 per index task at standard prices. It also reports planned API pricing of $2 per million input tokens and $10 per million output tokens initially, later $4 and $20. These figures are reported prices and estimates, not a calculation for an individual user’s workload.
Source: ThorstenMeyerAI.com
