Checked October 29, 2026.
Google announced Gemini 4 Carbon on October 28, emphasizing coding, cybersecurity and sustained multi-step work. Initial access is for selected cyber defenders, followed by trusted partners through Fairwind. Google has not given a general-release date. Google’s announcement.
Can you use it in Gemini today?
The public Gemini API catalog checked for this article lists Gemini 4 Argon, with no Carbon endpoint. Developers should wait for a documented model identifier rather than guess one. Gemini API model catalog.
The Gemini app’s release notes describe Argon’s October 13 rollout to Google AI Pro and Ultra subscribers. That entry does not establish Carbon access for those subscriptions. Gemini app release notes.
What do the numbers show?
Google’s comparison table gives Carbon 88.2% on SEC-bench Pro and 71.0% on CWE-bench, versus Argon’s 83.4% and 68.0%. These are benchmark scores, not a promise about your codebase. Google DeepMind’s comparison.
The methodology matters: Google computed SEC-bench results, while CWE-bench results came from Collinear’s evaluation. Settings and task selection affect comparisons. Evaluation methodology. Google also says the highlighted cybersecurity benchmarks were evaluated without guardrails. Announcement caveat.
Our take
For development teams, the useful next test is practical: can a model reproduce a bug, explain its cause, propose a small fix and pass regression tests? Keep human review and limited permissions in that workflow. A strong leaderboard result is a reason to evaluate a tool carefully, not to skip verification.