Google DeepMind is initially releasing Gemini 4 Carbon only to a select group of cyber defense professionals, as the company announced in the Google blog. Additional vetted partners will be added gradually via the Fairwind program, which is aimed at government agencies, operators of critical infrastructure, large technology platforms, and research labs.
As a result, Carbon is not available for selection anywhere in the Gemini app, including in Germany. According to Google, developers, businesses, and private users are to get the model as soon as possible; the company does not name a date.
What is Carbon supposed to do better than Argon?
Google describes Carbon as a model for long, autonomous work assignments, especially in programming. In Google's internal tests, the model kept working on a task for days, even after failed attempts. According to Google, before Carbon gets started, the model points out contradictions between sources instead of guessing. By Google's measurements, unsupported claims occur about half as often as with its predecessor Gemini 4 Argon, which Google introduced a good two weeks earlier.
The second focus is IT security. According to Google, Carbon achieves 88.2 percent in the SEC-bench Pro test for detecting vulnerabilities, 71.0 percent in CWE-bench for patching them, and 54.8 percent in ExploitGym for exploiting them. According to heise online, Argon reached 42 percent in ExploitGym. The staggered release takes place as part of a voluntary process in which the US government tests new models before launch.
The analysis company Artificial Analysis has already tested Carbon and ranks the model first out of 52 models in its Intelligence Index with 59 points. The median of comparable models is 42 points.
Gemini 4 Carbon in the measurement test
Independent values from Artificial Analysis, as of October 29, 2026, plus the API discount for recurring inputs according to Google
- 59points in the Intelligence Index, rank 1 out of 52 models
- 3×as many output tokens as the median: 200 instead of 64 million
- 100 swait time until the first token, versus 3.6 seconds for the median
- 95 %discount on inputs that stay the same across multiple requests
Carbon is not yet generally available; the measured values may change before the broad launch
Sources: Artificial Analysis, Google Blog
For the Index tasks, Carbon generated 200 million output tokens, more than three times as many as the median of 64 million. On average, around 100 seconds passed until the first token of a response appeared; the median is 3.6 seconds. Together with the regular API price, which is twice as high as the median, Carbon ends up only in 45th place out of 51 in cost per Index task.
What does Carbon cost?
For the API, Google is adopting Argon's introductory price at launch. After the introductory phase, the prices for input and output tokens double; Google leaves open how long the cheaper price will apply. Inputs that stay the same across multiple requests and are cached cost 95 percent less. Google lists API prices only in dollars.
In the Gemini app, you get the current top model Argon in the Google AI Pro and Ultra plans; the entry-level AI Plus plan does not include Argon. Which plan will include Carbon later is still open.
What this means for you
For your work in the Gemini app, nothing changes for now: the most powerful Gemini model you can sign up for remains Argon in the Google AI Pro plan. Signing up for a more expensive plan because of Carbon is therefore not yet worthwhile. According to heise online, with Argon there were around two weeks between the announcement and the release for paying users. The difference between free Gemini and Pro access is shown in the first lesson of the Gemini crash course.
If you plan to use Carbon via the API for your own tools or client projects, it is best to calculate with token consumption and not just the price per million tokens. A model that outputs more than three times as much text quickly eats up the advantage of a low token price. For a chat window with fast responses, a wait of 100 seconds until the first token is a poor fit.
If you work at a government agency or at an operator of critical infrastructure, your team can apply for access via the form on the Fairwind page. Google does not name a list of countries. How language models already help defend against attacks today is shown in the course GenAI for Cybersecurity.
Google introduced Gemini 4 Carbon on October 28, 2026.
Sources
