Google makes the case for deep-thinking models. On November 10 local time, Google officially launched Gemini 4 Carbon, positioning it as its strongest reasoning model to date, with a primary focus on complex software engineering, multi-step agentic tasks, scientific research, and cybersecurity.
Spend More Tokens, Think Deeper—Built to Solve Complex Problems
Compared with its sibling Gemini 4 Argon, Carbon is willing to spend more reasoning tokens on complex tasks for hypothesis exploration, result verification, and boundary checks, making it better at multi-step coding, complex debugging, and open-ended scientific research. It comes with a 1.04 million-token context window, an output limit of 65,000 tokens, Long Decode up to 1.04 million, and three levels of reasoning intensity. Paid developers can access it through the Gemini API and AI Studio, while enterprise customers and AI Ultra subscribers will also get access across Antigravity, the Gemini app, and AI Mode in Search.
Security and Research Both Score High
Cybersecurity is a major focus: 88.2% vulnerability discovery on SEC-bench Pro and 71.0% fix rate on CWE-bench; the Wiz team used it to uncover a zero-day vulnerability in less than two hours. On the research side, in CASP17 blind tests across seven low-homology protein families, it increased the average number of retrieved homologous sequences 40-fold. An even stronger proof point: on the Gray Swan injection benchmark, its attack success rate was 0%, making it one layer more hardened than Argon.
