article --summary
The summary
Google released Gemini 4 Carbon to paying developers, positioning it above the Argon model for demanding software engineering, cybersecurity, and scientific reasoning. The release also previews /swarm, a feature that coordinates several coding agents from one prompt, though Google has not explained how those agents share context or resolve conflicting changes.
Google says Carbon leads several benchmarks at under half the cost of rivals, but it has not offered enough detail to independently check those claims. Pricing matches Argon at launch and then doubles after November 30.
The article's most striking example is a formal verification project in which Google researchers used AlphaProof Code agents and Carbon to produce Lean proofs of security properties in Signal's cryptographic code, which Google says took weeks rather than years. Google has not said how complete the verification is, how much human work was involved, or whether the proofs will be published. The same capabilities raise misuse concerns, and Google's reported safeguard results come from specific simulated tests rather than proof of real-world resistance.
Editorial note: This is a summary of an article first published by The New Stack. The full piece is on their site.