New models
Gemini 4 Carbon: strong cyber ambitions, but no broad access yet
Google announces Gemini 4 Carbon for complex engineering and cybersecurity. Access remains limited; broad availability is not yet confirmed.
Google DeepMind announced Gemini 4 Carbon on 28 October 2026. The provider targets the model at long-running engineering tasks and cybersecurity. The announcement follows Gemini 4 Argon, but Carbon is not yet a generally available replacement that every developer can deploy immediately [1].
What Google reports about performance
According to Google, Carbon achieves 88.2% on SEC-bench Pro for vulnerability detection, 54.8% on ExploitGym and 71.0% on CWE-bench for repairs. Google also mentions a score of 59 and a first place on version 5.0 of the Artificial Analysis Intelligence Index. These are results reported by Google; we have not conducted these evaluations ourselves [1].
The provider also describes internal experiences with long-running coding tasks and generating secure code. Such examples can provide direction for an evaluation, but do not automatically predict quality on your own codebase. The announcement notes that the cited external cybersecurity benchmarks were evaluated without guardrails [1].
Phased access
Google is participating in the voluntary US pre-release testing process. From the announcement week, Carbon would become accessible to a selected group of cyber defenders; more trusted partners follow via the Fairwind Programme. Feedback and further safety measures precede a broader introduction for developers, businesses and consumers. A concrete date for that broad rollout is not in the source [1].
For the future introduction, Google mentions $2 per million input tokens and $10 per million output tokens. Cached input tokens would receive a 95% discount on the input price. This is announced pricing information, not proof that you can use the model today at those rates [1].
Significance for organisations
Our assessment: the limited access is the first planning point. Do not make a pilot project dependent on a yet-unconfirmed rollout date. Furthermore, define in advance which vulnerabilities, code changes and human checks an evaluation must include. A benchmark score does not replace a security review.
Read more
Sources
This article was produced with AI assistance from the original sources cited and checked against our editorial policy. Editorial policy