Gemini 4 Carbon is not heading straight to general availability. On October 28, Google announced plans to open the new AI model to selected cyber defenders that week. More trusted partners are to gain access gradually through the Fairwind Program.
Koray Kavukcuoglu describes Carbon as a model for demanding cyber work and engineering tasks lasting days. Google says it is participating in the US government’s voluntary pre-release testing process. It plans to gather feedback and refine safeguards before a broader release.
Google also points to fewer rule violations and unsupported claims in its own tests. These are vendor-reported results, not independently confirmed incident rates. Passing an evaluation does not establish that an agent will reliably stay within its task under different conditions.
The announcement establishes a limited rollout plan, not verified use by particular partners. The linked Fairwind page still names the earlier Argon model. The detailed safety report linked by Google also concerns Argon, not a separately reviewed Carbon report.