Google rolls out Gemini 4 Argon to security experts, not the public

Google has begun rolling out Gemini 4 Argon, its most advanced artificial intelligence model, but has not opened it to the public. For now, access is reserved for a select group of cybersecurity experts and organizations working on defense against digital attacks. The version currently in circulation has an output limit set at one million tokens. The company has stated its intention to gradually expand the user base, starting with paying users, following further checks.

The choice of a gradual rollout comes after the security incidents reported over the summer by several companies in the sector, including Google itself. The context in which Gemini 4 Argon debuts is dense with competing launches: the presentation follows by just a few days those of OpenAI’s GPT-6.1 Sol and Anthropic’s Claude Opus 5.5, a sign of a phase in which the major artificial intelligence labs are chasing closely-timed announcements.

Google has claimed top-tier results in several industry benchmark tests, including one that specifically measures security capabilities, in which the new model reportedly outperforms Astra, the system developed by OpenAI. It remains unclear, however, on what methodological basis this comparison rests, since the technical details of the tests have not been made fully public. Internally at the company, meanwhile, doubts circulate about the model’s actual performance in areas considered crucial for commercial adoption, such as programming: a sign that caution in the rollout concerns not only security risks but also the reliability of the claimed capabilities.

It should be noted that, at present, the news about the launch of Gemini 4 Argon comes from a single source, the statement released by Google; there is no independent confirmation of the technical details and benchmarks cited. The outlets that reported the news, while distinct from one another, rely on the same company announcement, and therefore do not constitute independent sources for the purposes of verifying the content.

The choice to start with cyber defense organizations, rather than general users or developers, points to a release strategy that favors sectors in which any model errors can be identified and corrected before a wider rollout. The one-million-token cap on output, imposed at this testing stage, remains for now the most concrete technical parameter disclosed about the version being distributed.