November 10, 2026 – Google has released Gemini 4 Carbon to paying developers, and it’s pairing the launch with a headline result: DeepMind says AI agents powered by the model helped formally verify key security properties of the code behind the Signal protocol.

Wider access, with limits

Carbon went to cybersecurity defenders two weeks ago through Google’s Fairwind Program. Google released it to paying developers on Tuesday, positioning it above Gemini 4 Argon for complex software engineering, cybersecurity and scientific reasoning. Access to Carbon’s advanced cybersecurity tools still requires approval through Fairwind.

Launch pricing is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after an introductory period. Google hasn’t said when that period ends for Carbon. Argon’s promotional rate runs through November 30, so developers should check Google’s pricing page before budgeting.

What DeepMind did with Signal

The work centers on AlphaProof Code, a new framework for formal verification, which means mathematically proving that code meets a written specification. Humans and AI models, including Carbon and Claude Opus 5.5, developed the specifications, and teams of agents running on Carbon then produced the proofs.

The agents wrote about 6.5 million lines of proof in Lean, a language whose compiler checks every step. DeepMind compiled libsignal to 64-bit Arm machine code and proved the properties against that compiled code, not just the design. The properties include forward secrecy, authentication, post-compromise security and handshake secrecy in Signal’s post-quantum handshake and Triple Ratchet.

To test the pipeline, the team reintroduced two previously fixed bugs, and the system caught both. The AI agents audited their own proofs and flagged about 1% for human review.

Why it matters, and what it doesn’t show

The work took four weeks for a team of five people and “quite a few tokens.” Comparable efforts have historically taken expert teams years, and the seL4 microkernel verification took 20 person-years. If the approach scales, formal verification could become practical for far more software.

It is not a blanket guarantee of Signal’s safety. The proofs cover selected properties, and they rest on about 26,000 lines of trusted specifications and machine-code rules that humans had to get right. Signal CTO Ehren Kret said earlier verification covered smaller portions of the code, and that DeepMind’s approach allows significant expansion. DeepMind is open-sourcing the verification artifacts and has published a technical paper, so outside experts can check the work.

Sources: Google, The New Stack, DeepMind paper coverage, TUN