Google releases Gemini 4 Argon, its first flagship model, prioritizing cybersecurity teams

Google officially released Argon, the first flagship model in the Gemini 4 series, on September 30 local time. Unlike a typical flagship model release, Argon has not yet entered the Gemini interface for ordinary users; Google first opened it through the controlled early access program Fairwind to a group of trusted cybersecurity defense organizations for testing.

According to Google, only after the safety guardrails are further improved will Argon gradually open to developers, enterprises and ordinary consumers, and the first users to gain broader access will include paid API customers and Google AI Ultra subscribers. This order itself constitutes a judgment: Google placed safety evaluation and pre-release access procedures before expanding the user base.

Google positions Argon as a flagship model for complex, long-process workflows, with key coverage of enterprise knowledge work such as software engineering, law and finance, as well as cybersecurity defense. Compared with the previous generation model, Argon's single output limit has increased from 64,000 tokens to 1 million tokens, meaning the model can continuously generate longer reasoning and execution trajectories in a single task, reducing reliance on manual mid-task intervention.

Google has already used Argon at scale internally, with thousands of employees using it to complete tasks such as specialized programming, deep research and writing. These are internal usage disclosures from Google alone, not independently verified results.

In a test by the quantum computing team, Google says Argon found a new optimization scheme within minutes, a 40% improvement over previously published baseline results. Another set of Argon agents was used to analyze performance telemetry data from Google data center clusters and automatically perform memory optimization; Google says that once the related schemes are fully deployed, they have already freed more than 300 TiB of memory, with an expected eventual saving of 500 TiB to 1 PiB.

Argon was also used in Google's internal large-scale C/C++ to Rust migration work, covering core open-source libraries such as re2 and libgav1 as well as the Fuchsia Zircon kernel, involving more than 800,000 lines of code. One specific case is the open-source video decoder libgav1: Google says an Argon agent rewrote about 32,000 lines of SIMD code in safe Rust, and the final version was 2.7 times faster than the previous Rust port while video output remained completely identical.