Google limits Gemini 4 Argon as it tests cybersecurity risks
Google is giving Gemini 4 Argon first to vetted cyber defenders, rather than releasing its most capable model publicly, because its ability to find and fix software flaws could also be misused. The model is built for long, complex work across coding and business tasks; Google has not set a date for wider access, leaving European developers and companies waiting to test it.
Bottom line — Google’s first wider access is planned for paid API customers and Google AI Ultra subscribers, but no release date has been announced.
Go deeper 5
-
Google says Argon can find, validate and patch critical software vulnerabilities autonomously; testers used it to uncover a flaw in hospital software that earlier advanced models had missed.
-
The company says the model can produce up to 1 million tokens in one output, up from 64,000 in previous Gemini models, supporting longer tasks without splitting them into repeated interactions.
-
Google reports Argon scored 68.9% on the Vals Index for professional work and 51.3% on AutomationBench. These are company-reported results, and other comparisons show Argon trailing rivals on some coding tests.
-
Google is monitoring the model’s reasoning and actions for behaviour that exceeds user intent, and says it is strengthening protections against misuse and prompt injection.
-
Google says broader access will start with paid API customers and Google AI Ultra subscribers. Introductory prices are €1.70 per million input tokens ($2) and €8.50 per million output tokens ($10); standard prices are €3.40 and €17 respectively.