On 30 September Google unveiled Gemini 4 Argon, billed as its most capable artificial intelligence model. There is a catch, though: at first only a small circle of partners in cybersecurity can use it. There is no date yet for the public.
Who can use it now
Access goes through Fairwind, Google’s security initiative. Its members include selected Google Cloud customers, government bodies and cybersecurity firms, as well as Google’s own internal teams. Anyone not on that list stays out for now.
Google justifies the decision with the risk of misuse. The model was trained with a particular focus on cyber defence and, according to the company, can “autonomously find, validate and fix” serious vulnerabilities in software. A capability like that is useful to those defending systems, but it can also be useful to those attacking them. Google also says it takes part in the voluntary US process that lets the government access advanced models before public release.
What it can do
Beyond security, Google points to programming and software engineering, and to complex professional work such as legal and financial tasks. Its employees already use it for debugging and code migrations. The company considers it comparable to leading models such as OpenAI’s Astra and Anthropic’s Opus in programming and cybersecurity tests.
The figures that have circulated so far come from Google’s own materials:
- DeepSWE v1.1: 77.9%
- CWE-bench v1: 68%, level with the best
- Output: up to 1 million tokens, against 64,000 for previous Gemini models
No independent lab has yet verified these results, so they should be read as figures claimed by the company.
Pricing and availability
For the API, the introductory price is $2 per million input tokens and $10 per million output tokens, with a 95% discount on input tokens already in the cache. Once the introductory phase ends it will rise to $4 and $20. Google has not announced prices in euros or Swiss francs.
For the wider rollout, the announced sequence is this: first paying API customers and Google AI Ultra subscribers, then everyone else. No date has been set for any of the stages, and so far there is no mention of either the free Gemini app or the AI Pro plan. Even an Ultra subscription does not give access today.
Who it’s for
For developers and ordinary readers, little changes in the immediate term: Argon is not available. It matters mainly to those who work in corporate security, and to those following the race between Google, OpenAI and Anthropic. If you already use Gemini in everyday apps, such as the notebooks for organising projects, you won’t see any change overnight.
What we think
The way the launch was handled shows how much security now weighs on model releases. Announcing a product and keeping it locked away is an unusual choice for Google, and it has a double effect: it protects against the most obvious misuse and at the same time puts the scores on show without letting others put them to the test.
That is the point to watch. A benchmark lead claimed by the maker, on a model almost nobody can try, is a promise and not yet proof. The real verdict will come when independent researchers and developers can compare Argon with its rivals on their real-world use cases. Until then it pays to stay cautious, and to keep a close eye on when and at what price it will reach Europe.
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
