Google releases Gemini 4 Carbon to developers and Ultra subscribers

Google made Gemini 4 Carbon broadly available on Tuesday, after a late-October release to a limited group of cybersecurity defenders through its Fairwind Program. Engadget reported that the earlier access went to select cybersecurity defense companies. Google calls Carbon its strongest model for cybersecurity and long-horizon agentic workflows, but says Gemini 4 Argon is the better choice for most software engineering and knowledge work. Carbon is built for tasks that test the limits of a model's intelligence, including complex, multi-step coding, debugging and open-ended scientific research.

Paying developers and enterprises can access Carbon via the API in Google AI Studio and Gemini for business, and it is also available in Android Studio for building. Carbon's introductory price is the same as Gemini 4 Argon's: $2 for every million input tokens and $10 for every million output tokens. Google AI Ultra subscribers, who pay $100 monthly according to Engadget, can use Carbon in Google Antigravity, the Gemini app and AI Mode in Search. Google is accepting applications to give 10,000 scientists free access for six months, with thousands already onboarded.

According to Google, Carbon achieves the top ranking on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, while costing less than half per task compared with the next best external model. On the new Multimodal Knowledge Agent benchmark created with Artificial Analysis, it ranks second while costing more than 6x less per task than the top model. It handles engineering work that needs multiple agents, including multi-package migrations, root-cause analysis across subsystems and redesigns of complex architectures.

Through the Fairwind Program, Carbon leads on SEC-bench Pro for finding vulnerabilities, ExploitGym for exploiting them and CWE-bench v1 for patching them, according to Google. In Google’s Project Fortify, the model achieved 100 percent recall on scanning and 95 percent precision on triage on an internal benchmark, and improved open-source recall by more than 25 percent relative to Argon. Wiz’s Scan for Good program used Carbon to find 0-day vulnerabilities in two major SaaS providers in under two hours.

Google also highlighted safety metrics. In simulated cyberattack testing, Carbon stopped 99 percent of attack paths, and the per-attempt breach rate dropped to 0.8 percent from 3.1 percent on Argon. Google reported that Carbon had a zero percent success rate for prompt injection attacks on the Gray Swan IPI benchmark, and it passed 92 percent of a 375-task agentic honesty benchmark, versus 70 percent for Argon.