Google opens Gemini 4 Carbon to cyber defenders first
DeepMind's newest frontier model leads the Artificial Analysis index and ships first to security teams, not the general public.
- #Gemini
- #cybersecurity
- #models
Google DeepMind announced Gemini 4 Carbon on October 28, 2026, and it is not a general consumer launch. Carbon is the company’s strongest model so far, aimed at cybersecurity work and long-running engineering tasks, and the first people who can use it are professional cyber defenders.
What Google is claiming
Carbon follows Gemini 4 Argon, which shipped about two weeks earlier. Google is positioning Carbon for work that lasts longer than a chat turn: hunting bugs across a large codebase, validating a fix, and carrying a task through many steps. Internal teams have reportedly kept Carbon jobs running for days.
On published benchmarks, coverage today lines up on four figures:
| Benchmark | Result | What it measures |
|---|---|---|
| Artificial Analysis Intelligence Index v5.0 | 59, ranked first | Broad model intelligence |
| SEC-bench Pro | 88.2% | Vulnerability discovery |
| ExploitGym | 54.8% | Exploit development |
| CWE-bench | 71.0% | Patching known weakness classes |
Those security scores are why the rollout is staged. A model that is strong at finding and developing exploits is also a model Google does not want in unrestricted public use on day one.
Who gets it, and what it costs
Access starts this week with a slice of cyber defenders. Wider access is supposed to move through the Fairwind Program to trusted partners, then to developers and businesses after more feedback and safety checks. There is no public release date. For most people, Gemini 4 Argon remains the newest widely available Gemini 4 model.
Introductory API pricing matches Argon: $2 per million input tokens and $10 per million output tokens, with cached input at a steep discount. Google has said those rates double after the introductory window.
Google also says Carbon is more willing than Argon to stop and surface uncertainty instead of inventing an answer. That claim is part of the safety story, not something outside reviewers have independently scored yet.
Why it matters
Frontier labs are now competing on agents that can stay on a software task for hours or days, not only on chatbot quality. Shipping the strongest model to defenders first is a product decision and a policy decision: the same capability that patches code can also write an exploit.
For engineering teams, the practical takeaway is narrower. Carbon is not something you can drop into a pipeline today unless you are in the early access group. Watch the Fairwind expansion, and treat the benchmark table as a vendor claim until independent replications land.
Sources
- Introducing Gemini 4 Carbon — Google, October 28, 2026
- Google announces Gemini 4 Carbon — 9to5Google
- Gemini 4 Carbon launch notes — explainx.ai