Google's most powerful AI model yet arrives under tight control
Broader access update, 14 November 2026
Google's November 10 rollout expanded Gemini 4 Carbon to paying API developers and enterprise users, with access also beginning for AI Ultra subscribers. This widened availability of an existing model; the more restricted cybersecurity release had begun roughly two weeks earlier.
Antigravity's /swarm preview coordinates groups of agents working across isolated workspaces. The rollout gives more teams an opportunity to test sustained coding and knowledge-work tasks involving several agents, rather than judging the model only from a short demonstration.
The introductory API rates are US$2 per million input tokens and US$10 per million output tokens. After November 30, the stated rates rise to US$4 and US$20 respectively. Google says demanding Carbon tasks can consume more tokens than Argon, so the lower headline token price alone does not establish which workflow costs less.
Teams assessing adoption should compare reliable completion of real tasks, human review time and total token usage. Those measures are needed to evaluate both cost and usefulness. The broader rollout announcement was published on November 10; the earlier restricted-access account is retained below as historical context.
Earlier restricted-access report
Just two weeks after Gemini 4 Argon was introduced, Google has announced Gemini 4 Carbon, which it calls its most powerful model yet, built for cybersecurity and long-horizon engineering tasks. Announcing it on 28 October, Google DeepMind's Koray Kavukcuoglu said the model tops the Artificial Analysis Intelligence Index with 59 points.
Benchmark numbers back the claim. Carbon scored 88.2 per cent on SEC-bench Pro, which tests whether a model can independently find real vulnerabilities buried in production software, 71 per cent on CWE-bench, and 54.8 per cent on ExploitGym, and it has climbed to fourth place on WebDev Arena for full-stack web development.
What sets Carbon apart from an ordinary model release is its purpose. This is not a chatbot upgrade. Google describes teams running it continuously for days to find and fix the root causes of subtle bugs in large codebases and low-level systems without retreating when they hit obstacles. It is designed as an engineering agent, not a conversationalist.
Internal results Google cites are striking: Carbon reportedly achieved more than double Argon's success rate at producing secure code, and the Chrome team says it generated high-quality patches ten times faster than another lab's frontier model. Wiz, Google's cloud security platform, notes that Carbon also leads the web-penetration benchmarks used in its own scanning programme.
Yet very few people can actually use it. From this week, access is limited to select cybersecurity professionals, with further trusted partners added gradually through Google's Fairwind Program. Google is taking part in the US government's voluntary pre-release testing process. Developers, enterprises and consumers must wait.
That restraint is the story. A model able to find and verify vulnerabilities at this level is equally valuable to defenders and attackers, and Google is explicitly testing that trade-off before opening the gates. Pricing tells the other half of the story: Carbon launches at Argon's introductory price - two dollars per million input tokens, ten per million output - production economics, not research pricing. The frontier is no longer just a leaderboard; it is a gate.
