Google opens Gemini 4 Carbon to paid developers and AI Ultra users
Google announced on November 10, 2026, that it is officially opening Gemini 4 Carbon to developers, enterprises, scientists, and consumers. For the previous two weeks, the model had only been available to some cybersecurity defenders through the Fairwind Program.
Paid-tier Gemini API users can access it in Google AI Studio and Android Studio; Google AI Ultra subscribers can use it in the Gemini app, Search AI Mode, and Antigravity. Pro and free users are not yet supported.
The API launch price is the same as Argon: $2 per million input tokens and $10 per million output tokens. The discount lasts until November 30, 2026, after which it rises to $4 per million input tokens and $20 per million output tokens. Google said Carbon is suited for tasks such as complex programming, deep debugging, and research that require repeated reasoning and verification with more tokens, while Argon is still recommended for most day-to-day work.
Google said that Wiz, its subsidiary, used Carbon in the Scan for Good project to scan two large SaaS providers, found a new zero-day vulnerability in less than two hours, and helped the relevant companies fix it, compressing research that would normally take weeks into one afternoon. This case used the Carbon for cyber version accessed through the Fairwind Program; the version now opened to ordinary users is the general-purpose version.
Original source:
Title: Gemini 4 Carbon is now available Publisher: Koray Kavukcuoglu (SVP, Google DeepMind and Chief AI Architect, Google) Publication date: 2026-11-10 URL: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/ Nature of source: official announcement Verification status: full text actually read in this report, verified on 2026-11-10; no access restrictions Supporting content: supports the core facts (the November 10 opening scope, the Fairwind-limited release two weeks earlier, API and Ultra access, pricing and end date, Argon positioning, the Wiz case, and the distinction of Carbon for cyber)
ABAB AI Insight
Koray Kavukcuoglu is Senior Vice President of Google DeepMind and Google’s Chief AI Architect, leading the development and product integration of the Gemini series of models. Gemini 4 Argon was first opened to cybersecurity defenders in the Fairwind Program on September 30, 2026, and was not broadly launched to paid developers and Pro/Ultra users until October 13. Carbon follows the same pattern: first released in a restricted manner to gather feedback and iterate on safeguards, then expanded to paid tiers.
The Fairwind Program is DeepMind’s controlled early-access channel for models with strong cybersecurity capabilities, prioritizing government, critical infrastructure, and core technology platforms, and requiring background checks and usage restrictions. Wiz, Google’s acquired cloud security company, tested Carbon through this channel in Scan for Good and reported finding zero-day vulnerabilities at two large SaaS providers. After the general-purpose version is opened, cybersecurity-specific capabilities still require a separate Fairwind application, reflecting Google’s tiered control over dual-use risks.
Compared with Argon, Carbon scores higher on independent benchmarks such as the Artificial Analysis Intelligence Index, but the official position is that it consumes more tokens for deep exploration, reasoning, and verification loops. In terms of pricing, the two launched at the same price and were raised in tandem after the promotional period ended. This design ties higher intelligence density to higher per-task cost rather than simple replacement.
This opening changes the immediate availability of frontier models for paid developers and Ultra users: complex multi-step programming and research tasks can now call Carbon directly, while day-to-day work is still steered toward the more efficient Argon. For enterprises, this means access to stronger long-horizon intelligence under controlled costs, but with the condition of higher token usage and continued restrictions on the cybersecurity-specific version.
ABAB News · Cognitive Law
In cybersecurity capabilities, restricted initial release followed by a general-purpose version is the fixed sequence under the Fairwind model.
Among models priced the same, the one that consumes more tokens is positioned as a tool for complex tasks rather than a full replacement.
Cases from Wiz and other internal or controlled partners can demonstrate capability, but do not mean ordinary users have obtained the same cybersecurity version.