Google unveils Gemini 4 Argon, flagship model for complex reasoning

Google DeepMind announced on September 30, 2026 the new flagship model Gemini 4 Argon. This is reported by Snow System and Tek.jp, both originating from Google DeepMind and from statements published on X by company executives, including Sundar Pichai.

The model is described as specialized in complex logical reasoning and long-duration workflows. One of the main novelties concerns the output token limit, expanded from 64,000 to one million: a leap that allows the system to handle more extensive tasks without intermediate interruptions, an element that the two sources indicate as the main technical advancement compared to the previous generation.

Initial distribution takes place through the Fairwind Program, aimed at a group of trusted testers, with declared priority to cyberdefense. This is therefore a selective access and not a general release to the public: Google has not indicated timelines for a wider rollout, nor do the consulted sources report a public availability date.

On the commercial front, the indicated launch price is $2 per million input tokens and $10 per million output tokens. Company executives, including Pichai himself, published details on X about internal benchmarks and the model’s design choices, without raw test data or a third-party verifiable methodology being made public.

The two sources consulted agree on the main features of the announcement — date, technical characteristics and distribution channel — despite coming from different outlets, one Japanese and one general-interest. Both, however, trace back to the same origin: communications and statements from Google DeepMind. This means that the benchmarks and performance claims for Gemini 4 Argon remain, for now, statements from the producing company, unconfirmed by independent analysis.