Google releases Gemini 4 Carbon, its most powerful AI model, for hard coding, research and cybersecurity work

Google on Tuesday released Gemini 4 Carbon, which it calls its strongest artificial intelligence model, to developers, businesses, scientists and its top-paying consumers, the company said in a Nov. 10 announcement by Koray Kavukcuoglu, Google’s chief AI architect.

An AI model is the underlying system that powers chatbots such as Google’s Gemini app.  Carbon sits above Gemini 4 Argon, a model Google released earlier.  Google says Argon remains the best choice for most coding and office work, while Carbon is meant for the hardest jobs:  long, multi-step programming and debugging, open-ended scientific research and finding security flaws in software.  Carbon spends more computing effort than Argon to think through a problem and double-check its work, Google says.

What Google says it can do

Google says Carbon ranked first on several outside tests of real work, including CursorBench 4.0, at less than half the cost per task of the next-best outside model.  Those are Google’s claims; the paper has not tested the model.

Google gave examples from its own researchers.  One team used Carbon with other tools to write millions of lines of computer proofs that check key security properties of the code behind the encrypted messaging app Signal, work Google says usually takes expert teams years.  Another team used it to find 40 times as many related protein sequences for hard-to-study proteins, which can help AI tools predict protein shapes.  Google said broader testing of that method is still underway.

Security risks and safeguards

Because a model that is good at finding software flaws could also help attackers, Google first gave Carbon’s full security abilities to vetted defenders through a program called Fairwind two weeks ago.  Google says its safeguards blocked 99% of simulated attack attempts by its own red team, a group that tries to break the system.  It also says Carbon is more honest than Argon, declining to make up missing experiment data 98% of the time in its tests.

Who can use it and what it costs

Paying developers and businesses can use Carbon now through Google AI Studio and Gemini for business.  Consumers can try it in the Gemini app and in AI Mode in Google Search only with a Google AI Ultra subscription.  Developers pay by the token, a small chunk of text.  The price is $2 per million input tokens and $10 per million output tokens until Nov. 30, then rises to $4 and $20.  Scientists can apply for six months of free access.

Source:  Google announcement, “Gemini 4 Carbon is now available,” Nov. 10, 2026.