Overview
- Google rolled out Gemini 4 Carbon on Tuesday to paying developers, enterprises and Google AI Ultra subscribers through the Gemini API, Google AI Studio, the Gemini app, AI Mode in Search and the Antigravity agent platform while keeping Gemini 4 Argon as the recommended model for most engineering tasks.
- Introductory pricing is $2 per million input tokens and $10 per million output tokens through November 30 and then doubles, and Google warns that Carbon’s deeper checking loops use more tokens so per-task costs can be higher than Argon despite the initial low per-token rates.
- DeepMind used an AlphaProof Code agentic workflow that paired human experts with Carbon to generate roughly 6.5 million lines of Lean proofs that verify key guarantees in Google’s build of libsignal, and Google has published the specifications and proof artifacts while noting exclusions such as side channels.
- Security partners report rapid vulnerability work using Carbon: Wiz found zero-day flaws at two major SaaS firms in under two hours and Google Cloud Vulnerability Research used Carbon to confirm and help patch two exploitable open-source bugs, but advanced exploit access remains gated to vetted Fairwind participants.
- Google and partners place Carbon at or near the top of several agent and coding benchmarks and report strong internal safety results, yet most performance and safety claims come from company or partner tests so independent benchmarking, real-world cost studies and adversarial resilience remain limited and worth watching.