NewsModels & ProductsNew
Google announces Gemini 4 Argon, first limiting access to vetted cyber defenders
Google announced Gemini 4 Argon on 30 September as its most powerful model, but initially released it only to trusted cybersecurity defenders; its benchmark scores are Google's own figures.
Evidence: 4 sources · 1 official record · 3 independent reports · 3 open questions
What happened
- On 30 September, Google announced Gemini 4 Argon, which it describes as its most powerful model yet.
- Access was first limited to trusted cyber defenders in Google's Fairwind Program, a limited-access program launched on 2 September.
- Google said broader access would come later, starting with paid API customers and Google AI Ultra subscribers.
- Google said it would offer a version without cyber guardrails to vetted defenders and its own internal teams.
Why it matters
A major lab holding back its flagship model from the public over cybersecurity risk is a sign of how seriously companies now treat offensive cyber capability in frontier models.
What is confirmed and what is claimed
Confirmed
Claimed, not independently verified
- Google says Argon delivers frontier performance on coding, enterprise knowledge work such as legal and finance, and cybersecurity defense. 1
- Google reports a score of 68% on CWE-bench v1, tied for first place, and 91.7% on the LVBench video benchmark. 1
- Google says Argon can produce up to 1 million output tokens. 1
Google's Gemini 4 Argon arrived one day after OpenAI's GPT-6.1 Sol, but with an unusual launch: the people who could use it first were vetted security teams, not the general public.
Context
Fairwind, the program Google set up in early September, is open to governments, Google Cloud customers, and selected cybersecurity partners. Participants must restrict use to staff working in areas such as incident response or penetration testing and must use controls such as multi-factor authentication. Google says it is prioritizing safety testing before a wider release.
The logic is that a model capable of finding and exploiting software flaws can help defenders patch them, but could also help attackers. Giving defenders a head start is one way labs are trying to manage that dual use. Google also said it would offer a version without cyber guardrails to vetted defenders, which shifts more of the safety burden onto access controls.
The benchmark figures in Google's announcement are self-reported. See our guide to reading benchmark claims, and our policy and safety map for how labs and governments are handling cyber risk.
What to watch
- A confirmed date for access through the paid API and consumer subscriptions.
- Independent evaluations, especially on cybersecurity tasks.
- Whether other labs adopt similar staged releases for their most capable models.
What we don’t know yet
- When general public access will begin. One secondary source reported a 13 October rollout to paid subscribers and the API, but other reports dispute this and we could not confirm it from Google.
- How the model performs on independent evaluations.
- What controls apply to the guardrail-free version beyond the requirements Google has described.
Evidence and sources
- Primary announcement Gemini 4 Argon: our next era of frontier intelligence
- Independent reporting Google releases Gemini 4 Argon, called its most powerful model yet
- Independent reporting Google unveils Gemini 4, long-awaited answer to OpenAI and Anthropic
- Independent reporting Google Launches Gemini 4 Argon With Guardrail-Free Access for Vetted Defenders