Google published a table of 17 benchmark results for Gemini 4 Carbon, set against Gemini 4 Argon, OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 and Claude Fable 5.1. Carbon comes first on 15 of them. Here is the whole table, what each test measures, and where Carbon falls short.
These are Google's numbers. Google ran most Carbon tests itself, with its highest thinking setting, and took the other models' scores from their makers' own reports or public leaderboards. Treat them as Google's claims until independent leaderboards catch up.
The full Gemini 4 Carbon benchmark table
Higher is better on every row. The best score in each row is highlighted. A dash means Google listed no score.
| Benchmark | Carbon | Argon | GPT-6 Astra | Opus 5.5 | Fable 5.1 |
|---|---|---|---|---|---|
| AA Intelligence Index v5.0 | 59 | 50 | 57 | 58 | 49 |
| Vals Index v2.2 | 70.4% | 69.2% | 65.0% | 70.2% | - |
| SWE-sweep | 9.6% | 4.3% | 9.5% | 3.6% | 2.9% |
| FrontierSWE 2 3rd | 61.2% | 55.0% | 65.5% | 62.3% | 56.3% |
| Terminal-Bench 4.0 (AA) | 65.2% | 57.1% | 59.1% | 59.6% | 52.0% |
| Agents' Last Exam v1.1 | 77.4% | 69.0% | 73.1% | 71.6% | - |
| OSWorld-2.1 (offline subset) 2nd | 81.5% | 74.7% | 80.9% | 87.2% | 76.8% |
| AutomationBench (Zapier) | 53.9% | 51.3% | 41.4% | 42.5% | 31.4% |
| Vending-Bench 2 (bank balance) | $16,484 | $13,718 | $15,515 | $9,235 | $5,422 |
| Chartography | 72.5% | 71.6% | 71.0% | 66.3% | 46.2% |
| Blueprint-Bench 2 | 58.6% | 54.4% | 49.7% | 51.2% | 41.9% |
| GraphWalks (256K to 1M) | 96.2% | 84.2% | 71.8% | 66.8% | 65.0% |
| RiemannBench | 80.0% | 76.0% | 72.0% | 69.6% | 65.6% |
| Terminal-Bench Science 0.1 (AA) | 64.8% | 52.9% | 63.3% | 59.0% | 43.3% |
| CWE-bench | 71.0% | 68.0% | 68.0% | 67.0% | 58.0% |
| ExploitGym | 54.8% | 42.0% | 42.4% | 34.5% | - |
| SEC-bench Pro | 88.2% | 83.4% | 85.4% | - | - |
Read from Google's announcement table, published October 28, 2026. Methodology: deepmind.google/models/evals-methodology/gemini-4-carbon.
Where Gemini 4 Carbon wins big
- Cybersecurity. ExploitGym, which checks whether a model can prove a bug is really exploitable, is the biggest gap in the table: 54.8% against 42.4% for the next model. Carbon also leads on finding bugs (SEC-bench Pro, 88.2%) and patching them (CWE-bench, 71.0%). More on our Carbon cybersecurity page.
- Very long documents. GraphWalks asks a model to follow connections through 256K to 1M tokens of text. Carbon scores 96.2%; the best non-Google model manages 71.8%.
- Business automation. On Zapier's AutomationBench, Carbon (53.9%) is more than 11 points ahead of GPT-6 Astra and Claude Opus 5.5.
- Running a business. Vending-Bench 2 hands a model a simulated vending business for a long stretch and checks the bank balance at the end. Carbon finishes with $16,484, the most of any model tested.
Where Gemini 4 Carbon doesn't win
Two rows go to other models, and two more are close enough to call a tie:
- FrontierSWE 2 (third place). On this tough software engineering test, GPT-6 Astra scores 65.5% and Claude Opus 5.5 scores 62.3%, ahead of Carbon's 61.2%.
- OSWorld-2.1 (second place). Operating a computer through screenshots, mouse and keyboard, Claude Opus 5.5 leads with 87.2% to Carbon's 81.5%.
- Near ties. On the Vals Index, Carbon's 70.4% is only 0.2 points above Opus 5.5. On SWE-sweep, its 9.6% edges GPT-6 Astra's 9.5%, and every model scores low on that very hard test.
So if your work is mostly driving a desktop computer, Claude Opus 5.5 still looks stronger on paper. For long coding agents, security work and huge documents, Carbon leads.
More results Google reported
In the Gray Swan test, attackers hide instructions inside documents and web pages to trick a model. With 15 tries per attack, none got through Carbon. Google's chart puts Claude Opus 5.5 and Claude Fable 5.1 at 1.0% and GPT-6 Astra at 8.5% on the same test.
Google also says Carbon is more honest with you: in its internal testing, unsupported claims fell by about half compared with Argon, doing more than asked ("overreach") fell 26%, and doing less than asked ("under-service") fell 29%.
What the tests measure, in plain English
| Group | Tests | What it checks |
|---|---|---|
| General | AA Index, Vals Index | Broad intelligence across many tasks, scored by Artificial Analysis and Vals AI |
| Agentic coding | SWE-sweep, FrontierSWE 2, Terminal-Bench 4.0 | Fixing real software issues and working in a terminal without help |
| Computer use | Agents' Last Exam, OSWorld-2.1 | Operating apps and a desktop to finish tasks |
| Knowledge work | AutomationBench, Vending-Bench 2 | Business workflows and running a simulated business |
| Multimodal | Chartography, Blueprint-Bench 2 | Reading charts and technical drawings |
| Long context | GraphWalks | Reasoning over 256K to 1M tokens of text |
| Science and math | RiemannBench, Terminal-Bench Science | Hard math and scientific work in a terminal |
| Cybersecurity | CWE-bench, ExploitGym, SEC-bench Pro | Patching, exploiting and finding security bugs |
Independent scores
One number in Google's table is already independent: Artificial Analysis ranks Carbon first on version 5.0 of its Intelligence Index, at 59, ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57). Other public leaderboards, like LMArena, haven't scored Carbon yet because it isn't publicly available. We'll add them here as they appear.
Benchmarks don't make videos.
Carbon writes text and code. For video and images, Google's Veo 3.1 and Nano Banana 2 are ready today on Lumeta.
Benchmark questions
Is Gemini 4 Carbon better than GPT-6 Astra?
On Google's table, yes on 16 of 17 tests. GPT-6 Astra wins FrontierSWE 2, a hard software engineering test, 65.5% to 61.2%.
Is Gemini 4 Carbon better than Claude Opus 5.5?
On 14 of the 16 tests where both have scores, yes. Claude Opus 5.5 leads on OSWorld-2.1, which measures operating a computer (87.2% to 81.5%), and on FrontierSWE 2 (62.3% to 61.2%).
Is Gemini 4 Carbon better than Gemini 4 Argon?
Yes, on all 17 tests in Google's table. The biggest jumps are in cybersecurity, long documents and science. See Carbon vs Argon.
What is Gemini 4 Carbon's Artificial Analysis score?
59 on the Artificial Analysis Intelligence Index v5.0, which ranks it first.
Sources: Google's announcement and benchmark table, evaluation methodology, 9to5Google.