On 28 October 2026, Google announced Gemini 4 Carbon, its most capable model so far and the successor to Gemini 4 Argon, which became generally available only two weeks earlier. Carbon is built for two kinds of demanding work: cybersecurity and long-horizon engineering tasks, such as projects that run for hours. It has taken first place on the independent Artificial Analysis Intelligence Index, but access is being rolled out in stages, starting with security teams.
How it compares
Artificial Analysis rebuilt its index in late October (version 5), with about 60% of the tests now private so labs cannot train on them. Scores are lower and more spread out than before. Here are the top models at their best setting:
Source: Artificial Analysis via BenchLeader, 3 November 2026. Best reasoning setting for each model.
The lead is narrow: one point is within the range where results can swap with a different test run. What stands out more is efficiency. According to Artificial Analysis, Carbon cost $7.76 per index task, about 23% less than Claude Opus 5.5 at $10.04, and used 29% fewer output tokens. It is still around six times the cost of GPT-6.1 Sol at maximum effort.
The security benchmarks
| Test | Gemini 4 Carbon | Others |
|---|---|---|
| CWE-bench, secure code (pass@1) | 71% | Argon, GPT-6 Astra and Grok 4.7 68%; Claude Opus 5.5 67% |
| SEC-bench Pro (finding real vulnerabilities) | 88.2% | Not on the public leaderboard yet |
| Gray Swan indirect prompt injection | 0.0% attack success | Argon 0.7% |
| AA-Omniscience (knowledge vs hallucination) | 54 | Claude Opus 5.5 48 |
Two caveats. Google ran the security tests without guardrails, and the released product will probably be a guarded version. And on AA-Omniscience, Carbon answers more questions correctly than Argon (58% against 48%), but it also makes things up more often when it is wrong: its hallucination rate rose from 4% to 11%. Higher scores do not mean you can skip checking its answers.
Who can use it, and when
Gemini 4 Carbon rollout
Announced
Pre-release testing with the US government's voluntary programme.
Cyber defendersNow
A subset of security teams gets access first.
Fairwind partnersUpcoming
Phased access for Google's 650+ trusted partners.
Developers and businessesUpcoming
"As soon as possible", after guardrails are refined. No date.
- Price
- $2 per million input tokens and $10 per million output at launch, with cached input 95% cheaper. The standard price will be $4 / $20.
- API
- No public model ID yet. Gemini 4 Argon remains the Gemini 4 model you can call today.
- Best for
- Security reviews, secure code generation and long, multi-step engineering tasks.
- Not best for
- Coding, where Claude Opus 5.5 still leads, and some knowledge-work tests, where Grok 4.7, Muse Wonder and Kimi K3.1 Pro score higher.
What it means for you
For most teams, nothing changes this month: you cannot call Carbon yet. In the meantime, Gemini 4 Argon is available at the same introductory price until 30 November, after which Argon rises to $4 / $20. If you run security reviews of code or data pipelines, ask your Google Cloud contact about the Fairwind Program. For everyone else, the more useful lesson is the pace: Google has shipped two frontier models in two weeks, so pick tools that let you switch models easily. Our monthly best AI models guide tracks the changes.
Check your understanding
1/3Who gets access to Gemini 4 Carbon first?
Access starts with security teams, then expands through the Fairwind Program.
2/3What was Carbon's score on the Artificial Analysis Intelligence Index v5?
It scored 59.4 at high effort, one point ahead of Claude Opus 5.5.
3/3What is a caveat on Carbon's AA-Omniscience result?
It answers more questions correctly, but when wrong it hallucinates more often: 11% against Argon's 4%.
You got 0 out of 3.
Read next: the best AI models in November 2026, Gemini 4 Argon, explained and GPT-6.1 Terra, explained.
Sources: Google's Gemini 4 Carbon announcement as reported by Constellation Research and Let's Data Science (28 October 2026); Tom's Hardware on Artificial Analysis results (3 November 2026); Artificial Analysis Intelligence Index v5 via BenchLeader (3 November 2026); Braindetox summary of Google's published security benchmarks. Security scores are Google's and were run without guardrails.
Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 3 November 2026. About me