The news of October 29thGoogle DeepMind Release New Generation FlagshipsModel Gemini 4 Carbon, about two weeks from Gemini 4 Argon. Of the most recent Inteligence Index v. 5.0 of the independent evaluation body Artificial Analysis, Carbon High ranked 59 at the top of 58 in Claude Opus 5.5 (Max)。
This upgrade focused on long-range programming and network security. Google indicated that the internal team had allowed carbon to run for several days to trace and repair hidden defects in large code libraries and bottom systems; the model was able to continue to try alternative solutions after having encountered a failed path。
In a web security assessment published in Google, Carbon obtained 88.2% on SEC-bench Pro, which discovered real software loopholes, 54.8% on ExpluitGym, and 71.0% on CWE-bench, which generated the root cause restoration programme. All three achievements were measured under conditions where no safety restrictions had been applied。
Carbon will be open to some of the cyber-security defence agencies from this week onwards and will gradually expand to more credible partners through Fairwind Programme. Google is participating in the voluntary testing of the United States Government ' s model prior to its release, and the general open hours for developers, businesses and ordinary consumers have not yet been published。
The initial pricing of the model was the same as that of Argon, with Token at $2 per million and the output at $10; at the end of the preferential period, respectively, it was $4 and $20。