Google rolls out Gemini 4 Argon, flagship model tested with US government

Google has announced Gemini 4 Argon, its new frontier model, releasing it first to a small group of trusted cyber defenders through the Fairwind Program. The company describes Argon as built for deep reasoning across complex, long-horizon workflows, real-world software engineering, enterprise knowledge work such as legal and finance, and cyber defense. The rollout is phased: Google is participating in the US government's voluntary early-access process for models and will expand access gradually while gathering feedback and updating guardrails before opening to developers, enterprises and consumers.

The model launches at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted 95% off the standard input rate. The most prominent technical change is the expansion of the output-token limit to 1 million tokens, a jump from 64 thousand in the previous generation. Google says that when the model has room to generate hundreds of thousands of tokens in a single pass, the added reasoning depth lets it solve hard problems in one go, without breaking the task into pieces.

Inside Google, Argon is already driving thousands of employees — Googlers — through specialized coding tasks, deep research and writing. The company's quantum-computing researchers report that the model improved the well-known space-time resource optimization baseline (qubits × gates) in bottleneck subroutines by 40%, in minutes. At the same time, Argon agents analyzed fleet-wide profiling telemetry and identified autonomous memory optimizations that freed more than 300 TiB (roughly 300 binary terabytes) in data-center memory, with a projected total saving of 500 TiB to 1 PiB (about 1,125 terabytes).

The most ambitious project Google discloses is large-scale migration of codebases from C/C++ to Rust. Argon agents handle core libraries such as re2 and libgav1 (tens of thousands of lines) up to the Zircon kernel of Fuchsia, more than 800 thousand lines. In libgav1, an open-source video decoder, the agents took an existing Rust port and replaced 32 thousand lines of SIMD code through many rounds of profile-guided experimentation, learning compiler output and writing memory-safe Rust that the compiler could auto-vectorize. The result: a memory-safe decoder running 2.7 times faster than the previous Rust port, with identical video output, and approaching the performance of optimized C++.

According to Google, the combination of coding, reasoning, multimodality and the ability to sustain long, multi-step tasks lets Argon excel across a range of enterprise workflows. The company's engineers use it for everything from daily debugging to algorithm design and massive migrations. But all the data comes from Google itself; there are no independent benchmarks, no third-party tests, and the model is still iterating on guardrails ahead of broad release. The voluntary process with the US government adds a layer of oversight, but also delays general availability.