What is Gemini 4 Carbon, and why is Google giving it to cyber defenders first?
Quotable summary
Gemini 4 Carbon is Google’s newest frontier AI model, announced 28 October 2026 and pitched on cybersecurity and long-running tasks. Instead of a public launch, Google is giving it first to “a subset of cyber defenders,” then to more trusted partners through its Fairwind Program. Developers, businesses and consumers get it later, “as soon as possible,” per Google. Translation: a model that’s very good at finding and exploiting bugs is going to the defenders first.
The facts
| Item | Detail |
|---|---|
| Announced | 28 Oct 2026 by Koray Kavukcuoglu, SVP of Google DeepMind and Google’s Chief AI Architect (Google blog) |
| Builds on | Gemini 4 Argon, which Google says launched two weeks earlier (same source) |
| Who gets it first | “A subset of cyber defenders” starting this week, then more trusted partners via Fairwind (same source) |
| Safety process | Google says it’s taking part in the U.S. government’s voluntary pre-release model testing process (same source) |
| Price | Introductory $2 per million input tokens / $10 per million output tokens, same as Argon, with cached input 95% off; $4 / $20 after the introductory period. Google hasn’t said when that period ends (same source; VentureBeat) |
What Google says it can do
These are Google-reported scores from the 28 Oct announcement, and Google says the security tests were run without guardrails.
| Benchmark | What it measures | Argon | Carbon |
|---|---|---|---|
| SEC-bench Pro | Finding real vulnerabilities | 83.4% | 88.2% |
| ExploitGym | Exploiting them | 42.0% | 54.8% |
| CWE-bench | Patching at the root cause | 68.0% | 71.0% |
Argon figures are from Google’s comparison chart as reported by VentureBeat.
Independent ranking: Artificial Analysis lists Carbon as the new #1 on its Intelligence Index v5.0, at 59, ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57), and marks it as having no publicly available API. Its testing also found a tradeoff: Carbon answers more factual questions correctly than Argon, but its hallucination rate on that test rose from 4% to 11% (VentureBeat).
Why the access gate?
- The same skills that find and patch bugs can also exploit them. That’s exactly what ExploitGym measures.
- It’s a pattern this month: Anthropic’s exploit-chain disclosures (our post) and OpenAI’s “Critical” cybersecurity rating for GPT-6.1 Terra (our post).
The Fairwind Program, briefly
- More than 650 partners globally (Google DeepMind Fairwind page).
- Priority groups: governments and national cyber authorities, critical-infrastructure operators and core technology platforms (Google, 2 Sep 2026).
- Rules: background checks on applicants, user-level authentication and phishing-resistant MFA, access only for internal security, incident-response or pen-testing teams, and no sharing, redistributing or selling access (Fairwind page).
What it means for you
- Regular users: nothing to do yet. You can’t use it.
- IT teams (our analysis): expect more AI-found CVEs, and plan to patch faster when they land.
FAQ
Can anyone use Gemini 4 Carbon? Not yet. Google says vetted cyber defenders get it first, then more trusted partners through Fairwind, then developers, businesses and consumers.
How is it different from Gemini 4 Argon? Google says it’s better at security and long-running engineering tasks, with more than 2x Argon’s success rate at generating secure code on Google’s internal benchmarks.
Are the scores verified? The three security scores are Google-reported. The #1 Intelligence Index ranking comes from Artificial Analysis, an independent evaluator.
How much will it cost? $2 per million input tokens and $10 per million output tokens at introductory pricing, rising to $4 and $20 afterward, per Google.