Google announced Gemini 4 Carbon, the new flagship after the Gemini 4 Argon, with an emphasis on cyber security and long-range agent missions. This week, the model will be given to selected cybersecurity experts, then access will be gradually opened to Fairwind program partners, and developers, businesses and ordinary users will be promised it "as soon as possible", without a date. The price is still like an organization: and #036;2 for 1 million input tokens and #036; 10 for a million weekends, later it will be twice as expensive. According to Google, Carbon spends days working on the problem and digging into the causes of subtle bugs in large code bases. In the table published by Google, the model is the first in the Artificial Analysis v5.0 index with 59 points and beat the GPT-6 Astra and Claude Opus 5.5 almost everywhere: inferior to them in FrontierSWE 2, and Opus also on OSWorld. On ExploitGym, where it is necessary to prove that the vulnerability can be exploited, it has 54.8% against 42% for an enterprise. Testers of the pre-release version wrote that in 3D tasks Carbon is noticeably stronger than Argon, but the front is weak, and it reacts, according to their estimates, 2-4 times slower. Channel @neuro.

Google announced Gemini 4 Carbon, the new flagship after the Gemini 4 Argon, with an emphasis on cyber security and long-range agent missions. This week, the model will be given to selected cybersecurity experts, then access will be gradually opened to Fairwind program partners, and developers, businesses and ordinary users will be promised it "as soon as possible", without a date. The price is still like an organization: and #036;2 for 1 million input tokens and #036; 10 for a million weekends, later it will be twice as expensive. According to Google, Carbon spends days working on the problem and digging into the causes of subtle bugs in large code bases. In the table published by Google, the model is the first in the Artificial Analysis v5.0 index with 59 points and beat the GPT-6 Astra and Claude Opus 5.5 almost everywhere: inferior to them in FrontierSWE 2, and Opus also on OSWorld. On ExploitGym, where it is necessary to prove that the vulnerability can be exploited, it has 54.8% against 42% for an enterprise. Testers of the pre-release version wrote that in 3D tasks Carbon is noticeably stronger than Argon, but the front is weak, and it reacts, according to their estimates, 2-4 times slower. Channel @neuro.