Google announced Gemini 4 Argon and is rolling it out first to cyber defenders through its Fairwind Program, before paid API and AI Ultra

Google DeepMind said on Tuesday that its new frontier model, Gemini 4 Argon, is now rolling out through the Fairwind Program to a set of cyber defenders it names as trusted testers, before it reaches paid API customers and Google AI Ultra subscribers. The company's own post carries the announcement and the numbers behind it.

Argon's output token limit is 1 million, up from the previous 64,000, and Google prices it at $2 per million input tokens and $10 per million output tokens, with cached input at 95 percent off. The company describes the model as built for long, multi-step work across software engineering, enterprise tasks like legal and finance, and cybersecurity defense.

The benchmark scores Google is citing

Google says Argon sets a new state of the art on DeepSWE v1.1, a real-world software engineering test, at 77.9 percent. On CWE-bench v1, which measures whether a model can remediate software vulnerabilities, Google reports a tie for first place at 68 percent. On AutomationBench, Zapier's benchmark for end-to-end business tasks, Google says Argon ranks first at 51.3 percent. On LVBench, a long-video understanding test, the company reports 91.7 percent. All four are vendor tests, run and reported by Google.

The post also names three internal uses. Google says Argon agents freed more than 300 TiB of memory across Google data centres in a fleet-wide optimisation, with 500 TiB to 1 PiB projected once the changes are fully rolled out. It says Argon agents replaced 32,000 lines of SIMD code in libgav1, Google's open-source video decoder, and the resulting Rust port runs 2.7 times faster than an earlier Rust version, with matching output. Argon agents are also working on migrating C and C++ codebases to Rust, up to 800,000 lines for the Fuchsia OS Zircon kernel. These are Google's own descriptions of its own internal work.

Cyber defenders get it first

The unusual part of the launch is the order. Google is releasing Argon to cyber defenders through the Fairwind Program without what the company calls its cyber guardrails, so those teams can use the model's full offensive-security capabilities to find and patch problems. Wiz is named in the post as an early user, through a free public-benefit programme called Scan for Good. Google says Argon uncovered a critical vulnerability in healthcare software used by hospitals, which previous frontier models had missed. That claim is Google's, retold from Wiz's work; no independent write-up is linked.

Google says Argon leads Gray Swan's Indirect Prompt Injection benchmark, which measures whether a model can be hijacked by hostile text a page or a document has been prepared with. It also describes internal monitors that watch Argon's chain of thought for signs the model is stepping outside a user's intent, and stop execution when they fire. Google says findings from those monitors are deliberately not fed back into training, to avoid teaching the model to hide its reasoning.

The general release date is not fixed. Google says Argon will reach paid API customers and Google AI Ultra subscribers as soon as the safeguards work is done, and points to the U.S. government's voluntary pre-release access process as one gate the launch passes through first.

Source

• Primary: Gemini 4 Argon: our next era of frontier intelligence, Google DeepMind, by Koray Kavukcuoglu, 2026-09-30.