Google's Gemini 4 Argon goes to 'trusted cyber defenders' first, at $2 in and $10 out per million tokens; no public date

Google announced Gemini 4 Argon on 30 September in a post by Koray Kavukcuoglu, SVP of Google DeepMind and Google's chief AI architect. It is rolling out first to a set of cyber defenders through what Google calls its Fairwind Program; everyone else waits.

Who gets it, and when

Google says releasing capabilities at this level "requires a phased approach". It is taking part in the US government's voluntary process for pre-release model access and will widen access gradually, gathering feedback from early testers "as we iterate on guardrails" before release to developers, enterprises and consumers "as soon as possible", starting with paid API customers and Google AI Ultra subscribers. No date is given.

The introductory price is $2 per million input tokens and $10 per million output tokens, with cached input at 95% off the input price. The output limit rises to 1 million tokens from 64,000.

What Google says it has done inside Google

Thousands of Googlers are already using it, and the post gives three examples. In quantum computing research, Argon beat a published baseline for optimising a subroutine's spacetime resources by 40% "in a matter of minutes". A team of Argon agents analysed fleet-wide profiling data and applied memory optimisations across Google's data centres, freeing over 300 TiB once rolled out, with 500 TiB to 1 PiB of total savings estimated. Argon agents are migrating C and C++ codebases to Rust, from core libraries up to the 800,000-plus-line Fuchsia Zircon kernel, with the rewrites under automated and manual audit before production. For the libgav1 video decoder, agents replaced 32,000 lines of SIMD code with safe Rust that the compiler vectorises itself; Google says the result runs 2.7x faster than the earlier Rust port with identical output.

Benchmarks Google cites

DeepSWE v1.1 at 77.9%, which Google calls a new state of the art for long-horizon software engineering. First on the Vals Index, which weights finance, coding, legal and tax work by their share of US GDP. First on Zapier's AutomationBench at 51.3%. LVBench, for long video understanding, 91.7%. On CWE-bench v1, for remediating security vulnerabilities, a tie for first at 68%. All are Google's own reported figures.

Cyber defence without guardrails

Google trained Argon to find, validate and patch vulnerabilities autonomously, and will release it "without cyber guardrails" to trusted defenders and its own teams. Wiz is using it through its Scan for Good programme; Google says it found a critical vulnerability exposing personal information in healthcare software used by hospitals worldwide, "a severe risk that previous frontier models had missed".

The safeguards it names

Four areas: refusing cyber and CBRN misuse while preserving legitimate dual-use research, with monitoring of the model's internal activations and red-teaming of the safeguards; resistance to indirect prompt injection, where Google says Argon leads on Gray Swan's benchmark; misalignment mitigations that monitor Argon's chain of thought and actions and stop execution when necessary; and sealed sandboxes before high-risk training or evaluation.

On monitoring, Google says it used a similar system during training runs, alerting an incident response team, and took "careful precautions against feeding the findings back into training" so as not to shape Argon's reasoning to evade monitoring. It urges the rest of the industry "to preserve reasoning transparency".