Google's new Gemini 4 Carbon AI tops rivals in coding and security tests

Google has announced Gemini 4 Carbon, a new AI model it says leads rivals in coding and cybersecurity. It arrives only two weeks after Gemini 4 Argon. For now, only select security companies can use it.

What Google claims

On the AA Index v5.0, a general test of AI ability, Carbon scored 59. Google says that is just ahead of Claude Opus 5.5 and GPT-6 Astra. It also claims a near sweep in areas such as agentic coding (AI that carries out coding tasks on its own), computer use, knowledge work, and science and math.

Security is the big focus. Google says Carbon is better at finding and ranking software weaknesses, and at fixing them at the root cause. It also says Carbon is twice as likely as Argon to write secure code. Google adds that it resists prompt injection, which is when hidden instructions try to trick an AI.

Who gets it and when

• First: select cybersecurity defense companies, for pre-release testing.
• Next: more trusted partners through Google's Fairwind Program.
• Last: a wide release to everyone.

Google followed the same path with Argon, which is now in the Gemini app and AI search mode. The material does not give a price or a date for Carbon's wide release.

What to keep in mind

The scores come from Google's own benchmarks. Independent testing has not been reported yet. Everyday users should expect to wait before trying Carbon, and should treat the claims as promises until others confirm them.