Google unveiled Gemini 4 Carbon on 28 October, describing it as its next frontier model. It arrives barely two weeks after Gemini 4 Argon and focuses above all on cybersecurity and long tasks that take many steps. As with Argon, though, access is limited to a very small group: you won’t be able to try it any time soon.
What changes compared with Argon
The model is designed to find vulnerabilities in software, exploit them in controlled tests and write the fixes. Google reports these results:
- SEC-bench Pro (vulnerability discovery): 88.2%, billed as a new state of the art;
- ExploitGym (exploiting flaws): 54.8%, also a claimed record;
- CWE-bench (fixing): 71.0%;
- WebDev Arena: 1768 Elo points, up 89 on before, fourth place;
- Artificial Analysis Intelligence Index v5.0: a score of 59, first in the ranking.
In internal tests, according to the company, the ability to generate secure code has more than doubled compared with Argon. These are figures published by Google itself: until independent researchers can measure the model in real-world conditions, they should be read as the manufacturer’s claims.
Google also devotes space to the model’s safety. It says attempts to “cheat” on coding tests have all but disappeared and that indirect prompt injection attacks never succeed in the benchmark used. The safeguards are the same as for Argon.
Who can use it
From this week Carbon is available to a subset of cyber defenders. Access will then widen to other trusted partners through the Fairwind programme, the same channel already used for Argon. For developers, businesses and ordinary users, Google speaks of a release “as soon as possible”, without setting a date. The company also takes part in the US government’s voluntary pre-release testing process.
The announcement names no specific countries or regions. For Italy and Switzerland, then, there is nothing to say about availability: no announcement, no app to update.
Pricing
For the API, Google quotes an introductory price of 2 dollars per million input tokens and 10 dollars per million output tokens. Once the introductory period ends, it rises to 4 and 20 dollars. Input tokens already in the cache cost 95% less. There are no prices in euros or Swiss francs, and it has not been said whether a consumer subscription including Carbon will exist.
Who should care
For the ordinary reader, today, the news is mostly indirect. A model that finds and fixes flaws better than its predecessors can make the browsers, apps and services we use every day more secure: Google cites the Chrome team, which reportedly obtained fixes of far higher quality than those from another leading model. If you work at a company with a security team, or you develop software, you would do well to follow the Fairwind programme.
What we think
Launching to small groups first is becoming a pattern: defenders first, everyone else later. It is an understandable choice, because a model that is good at exploiting vulnerabilities is useful to attackers too, and giving defenders a head start makes sense. It does come at a cost in transparency, though: the best results remain unverifiable by third parties, and the Artificial Analysis ranking measures what the company has been able to show.
Two things to watch over the coming days. The first is whether the general release really arrives quickly or remains a promise, as with Argon. The second is the final price: the doubling after the introductory phase will decide whether Carbon is a practical choice for small teams or only for those with a sizeable budget. For now, nothing changes for anyone who uses Gemini every day: the models already available are the same as before.
Written by the SpazioiTech newsroom with the help of AI tools, from facts verified across at least two independent outlets, and checked before publication. Spotted a mistake? Tell us.
Translated from the Italian original. Read in Italian
