Google has a new flagship AI model, and for once it isn't rushing it to your phone or browser first.
Gemini 4 Carbon, announced Wednesday, is the company's second frontier model release this month, following Gemini 4 Argon, which only finished its own rollout to everyday users two weeks ago. Google says Carbon improves on coding, agentic tasks, scientific reasoning, full-stack web development and multimodal understanding, and that it currently sits at the top of version 5.0 of the Artificial Analysis Intelligence Index, a third-party benchmark that ranks large language models against each other.
But the headline feature here is security, not speed.
According to Google's own benchmarks, Carbon more than doubles the success rate for generating secure code compared with Argon. The company also claims Carbon performed more than 10 times better than what it describes only as "another lab's large frontier model" with writing high-quality software patches. Google says Carbon held up well on three external benchmarks built specifically for security work: SEC-bench Pro, which tests a model's ability to find vulnerabilities, ExploitGym, which tests exploitation, and CWE-bench, which tests patching. None of those figures have been independently verified.
Carbon runs on the same safety and security stack Google built for Argon, but the company says it pushed further on alignment, the industry term for making a model's behavior match what its developers actually intend.
That security focus is also shaping how Carbon reaches people. Rather than pushing it straight to consumers, Google is following the US government's voluntary AI review process, a framework under which companies share frontier models with federal reviewers before wide release. Practically, that means Carbon isn't going to the public this week at all. Instead, Google plans to give access to a subset of cyber defenders, security teams whose job is finding and closing holes before attackers do.
A public rollout will come later, though Google hasn't said when. When it lands, developers using Carbon through Google's API will pay $2 per million input tokens and $10 per million output tokens, tokens being the chunks of text a model reads and generates. Cached input tokens, meaning text the model has already processed recently and doesn't need to reprocess from scratch, will cost 95 percent less than fresh input.
The pricing lines up with what Google charged for Argon, which suggests Carbon is meant to sit alongside it rather than replace it outright, at least for now. Whether Carbon eventually folds into Gemini's consumer apps the way Argon did will depend on how that defender rollout and government review play out, and Google hasn't put a date on either.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.