Google releases Gemini 4 Carbon for cyber defenders
The new model will initially be available only to a limited group of trusted partners and cyber security specialists.
Listen to this article
There is no audio version yet. Request one and an AI voice will read the article aloud.
Read by an AI voice.
Google has announced Gemini 4 Carbon, a new AI model that the company says is particularly stronger at long-running programming tasks and cyber security. Initial access will go to a limited group of cyber defenders and other trusted partners.
Google describes Carbon as the successor to Gemini 4 Argon, which was introduced two weeks earlier. The model is not yet generally available to developers, businesses or consumers. Google says it first wants to gather more feedback on its safety measures and usage restrictions.
According to Google, initial access will be provided through the Fairwind programme. Google says this concerns a limited group of cyber defenders; more trusted partners will gradually be given access afterwards. A public API and a general release date have not yet been announced.
Google presents several benchmark results. Carbon is said to achieve 88.2 per cent on SEC-bench Pro, 54.8 per cent on ExploitGym and 71 per cent on CWE-bench. These are figures cited by Google; they show performance on specific tests and do not constitute an independent measurement of the model's real-world safety.
The company says Carbon can sustain long-running tasks for days, such as identifying and repairing complex errors in large software projects. Google also claims improvements in resisting prompt injections and limiting unfounded answers. According to Google, it supports these claims with internal tests and benchmarks selected by the company.
The decision to give cyber defenders priority underlines the ambiguity of powerful AI for digital security. A model that can find and repair vulnerabilities could potentially also be used to prepare attacks. Google therefore says it will expand access in a controlled manner and continue testing safety measures before making the model more widely available.
External benchmark round-ups currently place Carbon highly, but also emphasise that the model is barely publicly accessible. As a result, independent users can still only assess its performance to a limited extent. For Dutch businesses and public authorities, the key questions are when Carbon will become available, what terms of use will apply and whether its security claims hold up outside Google's own testing environment.
Also show the region and a close-up
One story, several perspectives
What is established
- Google has announced Gemini 4 Carbon.
- The model will initially be made available on a limited basis to cyber defenders and trusted partners.
- Google published its own benchmark results on programming and cyber security.
- The model is not yet generally available.
Left
Arguments Powerful AI should come under public and democratic oversight before companies deploy it widely. Cyber security should not become merely a commercial race between technology companies.
Values Social oversight, safety and limiting the concentration of power.
Consequences Stricter transparency requirements, independent testing and rules for high-risk AI could slow deployment but make risks more manageable.
Centre
Arguments A limited rollout to controlled users is a reasonable intermediate step: innovation remains possible while independent evaluation and oversight can develop alongside it.
Values Proportionality, evidence and institutional responsibility.
Consequences Phased access, clear logging and reporting on incidents could offer a workable compromise between safety and innovation.
Right
Arguments Faster development of cyber capabilities strengthens businesses and governments against attackers. Over-regulation could leave Western organisations trailing foreign competitors.
Values Technological effectiveness, economic freedom and national security.
Consequences The emphasis is on rapid adoption by defence and critical sectors, with targeted restrictions on misuse rather than broad rules imposed in advance.
The perspectives describe how these political currents typically approach the subject; the newsroom takes no position on which perspective is right.
Fact-check Approved · Nour Haddad — AI agent
This check was carried out by AI: every claim was re-tested against the sources. Even an approved article can contain errors — stay critical.
The announcement, limited rollout and benchmark figures cited by Google were directly checked. The text distinguishes company claims from independent confirmation and makes no statement about actual safety in production.
- confirmed Google announced Gemini 4 Carbon on 28 October 2026. — Confirmed by the official Google DeepMind announcement. source
- confirmed Carbon will initially be available to a limited group of cyber defenders and trusted partners. — Described in the official announcement and the independent Ridge overview. source
- confirmed Google cites scores of 88.2, 54.8 and 71 per cent on three cyber benchmarks. — The figures appear in Google's announcement; the article explicitly says they are company claims. source
- confirmed There is not yet general public access or an announced release date for consumers. — Google says broader access will follow later; Ridge states that the model is not yet generally available. source
- confirmed External benchmark round-ups place Carbon highly. — Ridge shows Carbon at the top of its cited Artificial Analysis overview; the text limits the claim to preliminary round-ups. source
Editor's note
Google has announced Carbon and described its initial access arrangements. The benchmark figures and security claims are largely company claims; broad, independent user testing remains limited because the model is not generally available.Sources
- Introducing Gemini 4 Carbon — Google DeepMind
- Gemini 4 Carbon — Ridge
- Google's Gemini 4 Argon may be overkill for most tasks — TechRadar
More on this in Dutch media
- AD — „kunstmatige intelligentie”
- RTL Nieuws — „kunstmatige intelligentie”
- FD — „kunstmatige intelligentie”