Carbon starts with limited access

Google’s October 28 announcement places Gemini 4 Carbon in sustained engineering work and cyberdefense: finding vulnerabilities, validating them and fixing their causes. Access begins that week with a subset of cyber defenders, followed gradually by additional trusted partners through Fairwind. The company says it is gathering feedback and refining safeguards before broader availability for developers, businesses and consumers. The announcement therefore describes a phased rollout.

The evaluation document shows why test settings matter. For Agents’ Last Exam v1.1, Google used 122 tasks, a five-hour window and enabled safety filters. It used the October 17 task version without the October 26 update. Filtered responses became empty strings, although the task episode could continue. For OSWorld 2.1, the reported result covers only the 82-task offline subset: Google took the best of three runs, with one attempt per run. Unless otherwise specified, competitor results come from their providers; this is not a uniform independent comparison.

MarkTechPost’s report organizes the release around capabilities, safety changes and access. Its benchmark discussion attributes most competing results to provider reporting. That secondary account helps locate the announcement’s themes, but does not supply an independent test of the model.