Google Launches Gemini 4 Carbon for Finding Vulnerabilities

Google has released Gemini 4 Carbon, its second flagship model this month. This one is built for cybersecurity. The model finds vulnerabilities in code with 88% accuracy. It fixes the bugs it finds 71% of the time. Google says the share of secure code the model generates has more than doubled compared to its previous model, Argon. Argon came out two weeks earlier. On the updated Artificial Analysis Intelligence Index, which ranks AI models, Carbon now sits in first place with a score of 59. Pricing stayed the same as Argon: $2 per million input tokens and $10 per million output tokens. Tokens are the small chunks of text that AI models process.

A focus on code security is exactly what agentic AI models tend to lack. They write code fast, but they rarely catch their own vulnerabilities. For agentic coding work overall, Opus and Fable, Anthropic's code-focused models, remain the stronger choice. Gemini's edge, for now, shows up mostly in narrow benchmark records. But when an AI agent writes code without a human checking it, that is a real risk. Running a separate pass with a security-focused model like Carbon is a sensible precaution. It is not just hype.