Gemini 4 Argon targets long coding jobs with a million-token output limit

Google released Gemini 4 Argon on September 30, its new frontier model aimed at long, multi-step professional work rather than quick chat. The headline change for engineers is the output budget: Argon can produce up to 1 million tokens in a single response, a large jump from the previous 64K ceiling, which matters when a model has to rewrite a whole file or emit a long migration in one pass. In its announcement, Google reports 77.9% on DeepSWE v1.1, a benchmark for real long-horizon coding tasks, and backs it with internal results: codebase migrations from C and C++ to Rust that run past 800,000 lines, a 2.7x speedup on the libgav1 video decoder, and a 40% gain over published baselines on a quantum optimization problem.

The other theme is security. Argon ties for first at 68% on CWE-bench v1 for fixing vulnerabilities, and Google is releasing it first through a "Fairwind Program" for trusted cyber defenders before any general rollout, an unusual gate that signals concern about dual use. Pricing starts at an introductory $2 per million input tokens and $10 per million output, rising later to $4 and $20, with a 95% discount on cached input. Broader access is promised for paid API customers and Google AI Ultra subscribers, without a firm date.

Why it matters

If you build coding agents or evaluate models for engineering work, the million-token output and the DeepSWE score are the numbers to test against your own tasks, and the defender-first release tells you Google expects this model to be capable enough to misuse.