Two weeks ago, Google introduced Gemini 4 Carbon — its strongest model — and kept it behind a velvet rope, open only to a hand-picked set of cybersecurity defenders in its Fairwind Program. On Tuesday, the rope came down. Carbon is now available to paying developers through the Gemini API, to enterprises through Gemini for business, and to Google AI Ultra subscribers in the Gemini app, AI Mode in Search, and Google Antigravity.

Google's numbers, on its own terms, are striking. Carbon ranks first on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench, at what Google claims is more than twice the lower cost per task than the next-best external model. On the security side, where Carbon was born, it found zero-day vulnerabilities in two major SaaS providers in under two hours during Wiz's Scan for Good program — work Google says would normally take weeks. Argon stays the default for most engineering work; Carbon is for the jobs where the answer has to be right.

Pricing is the part worth reading twice. Carbon costs $2 per million input tokens and $10 per million output tokens through November 30 — the same introductory rate as Argon. On December 1, both double to $4 and $20. Treat the next three weeks as a trial window, because the economics change on the first of the month.

The more interesting announcement may be tucked inside the launch: Google is previewing `/swarm` in Antigravity, which turns a single prompt into a coordinated team of agents you can watch in real time. Carbon takes the roles that demand the deepest reasoning — architectural design, coordination, auditing — while faster, lighter models work in parallel, with adversarial verifiers testing every change. It's the clearest sign yet that Google thinks the next frontier isn't a single smarter model, but a managed team of them.

There's a safety argument underneath all of this: a model this good at finding vulnerabilities could be pointed the other way. Google says Carbon blocked 99% of attack trajectories in simulated cyberattack testing and recorded a 0% prompt-injection success rate on Gray Swan's benchmark.

The same model did the work behind this week's other headline: the AlphaProof Code framework that formally verified Signal's encryption down to the machine code. That was Carbon in a research setting. Now it's Carbon as a product, priced by the token, with a deadline on the discount.

— via Google Blog