Google launches frontier model Gemini 4 Carbon

Google has begun rolling out Gemini 4 Carbon to developers, enterprises, scientists and consumers, positioning it as its strongest model yet for long, multi-step agentic work — deep coding and debugging, open-ended scientific research, and cybersecurity — while its sibling model, Gemini 4 Argon, released a few weeks earlier, remains Google’s recommendation for most everyday software and knowledge work. Carbon uses more compute than Argon to reason more thoroughly and run longer verification loops, and Google had already quietly released it to cybersecurity defenders two weeks earlier through its Fairwind Program.

Google says Carbon ranks first on three independent benchmarks — CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench — at more than double the cost efficiency of the next-best outside model. Pricing matches Argon’s introductory rate: $2 per million input tokens, and, separately, $10 per million output tokens; it’s available now through the API inside both Google AI Studio and the Gemini for business console. Google AI Ultra subscribers also get it inside the Gemini app, AI Mode in Search, and a new Google Antigravity feature called /swarm that turns one prompt into a coordinated, trackable team of agents.

The company is leaning hardest on Carbon’s cybersecurity claims. Through Fairwind, it says Carbon leads benchmarks for finding, exploiting and patching vulnerabilities, and that Google’s own Project Fortify system hit 100% recall and 95% precision scanning internal code, with external testers at Wiz using it to surface zero-day flaws in major SaaS platforms within hours rather than weeks. Google also reports its safety testing blocked 99% of simulated attack paths, cut the per-attempt breach rate to 0.8% from 3.1% on Argon, and — for the first time on this benchmark — drove a prompt-injection attack success rate to zero on Gray Swan’s test, while cautioning that prompt-injection defenses are still an area it is working to improve.

Google also points to gains on an honesty benchmark it runs internally, where Carbon passed 92% of 375 tasks testing whether a model will decline to fabricate missing data or flag flawed methodology, up from 70% for Argon. In outside evaluations, Carbon ranked #2 on a new multimodal “knowledge agent” benchmark built with Artificial Analysis at a fraction of the top performer’s cost, and placed at or near the top of financial-research benchmarks run by Hebbia and Balyasny Asset Management.

Why it matters

Google is explicitly pricing and positioning Carbon as a premium, narrower tool rather than a blanket upgrade — reserving it for the hardest coding, research and security work while keeping Argon as the default — a split that signals the company expects frontier-model costs and capability gaps to keep widening rather than converging.