Google releases Gemini 4 Carbon
Google released Gemini 4 Carbon, focused on cybersecurity and long-horizon tasks; it can work autonomously for days on complex multi-step engineering problems. Versus Argon, secure code generation success more than doubled, with major drops in alignment and hallucination issues. Launched at introductory pricing, first open to select cyber defenders and trusted partners.
Google Releases Gemini 4 Carbon
Gemini 4 Carbon is Google's most powerful frontier model to date, focused on cybersecurity and long-horizon task capabilities.
It can work autonomously for days on complex multi-step engineering problems (such as locating and fixing vulnerabilities in large codebases and low-level system debugging), with clear improvements in coding, agentic tasks, scientific tasks, full-stack web development, and multimodal understanding. Compared with the previous generation, Argon, its success rate in secure code generation has more than doubled, and alignment and hallucination issues have dropped significantly (for example, unsupported claims are reduced by about 2x, and high-severity misalignment alerts are down nearly 80%). It leads on multiple cybersecurity benchmarks (such as SEC-bench Pro at 88.2%, vulnerability detection and patch generation, etc.).
It is currently being launched at introductory pricing (input $2/million tokens, output $10/million tokens, the same as Argon, and will double later). It is first gradually opening to select cybersecurity defenders and trusted partners (including pre-release testing with the U.S. government), and will later be made more broadly available to developers, enterprises, and consumers.
Source: Google Gemini official launch
