Google unveils Gemini 4 Carbon, an upgraded AI model focused on coding and cybersecurity

Google introduced Gemini 4 Carbon, the latest iteration of its Gemini 4 series, positioned as an improvement over the recently released Argon model. The company says Carbon delivers notable gains in coding, agentic tasks and scientific analysis, and it is particularly strong in cybersecurity applications, claiming a "significant leap" in vulnerability detection, validation and remediation. According to Google, the model is twice as likely to produce code without vulnerabilities compared with Argon, and the Chrome team reports it is ten times more effective at generating high-quality patches than non-Google models. Carbon also achieved an 89-point increase on the WebDev Arena benchmark, placing it fourth on that leaderboard.

In internal testing, Carbon outperformed competing frontier models such as GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 on most metrics, though the performance gaps were relatively small. The model also recorded a 0 percent success rate for prompt-injection attacks in Google's Grey Swan evaluation, improving on Argon's 0.7 percent rate.

Google said the rollout will follow a deliberative approach similar to Argon, involving participation in the voluntary testing framework set by the U.S. government and eventual deployment through its Fairwinds cyber defense program. Initial access is expected for AI Ultra subscribers and developers using the Gemini API, with a temporary discounted API pricing of $2 per million input tokens and $10 per million output tokens. Google did not specify when the introductory pricing will end.

"Carbon shows a significant leap in vulnerability detection, validation, and remediation," a Google spokesperson said.

"In the Grey Swan evaluation, Carbon resisted every attempted prompt-injection attack," the company added.