Google Opens Gemini 4 Argon First To Vetted Cyber Defenders

Google has started giving vetted cybersecurity defenders access to Gemini 4 Argon, SiliconANGLE reports, placing the new frontier model inside a restricted channel before a general developer launch.

The first outside users are in Fairwind, Google’s controlled cyber access program.

Fairwind began on Sept. 3 with Gemini 3.8 Flash Cyber, a smaller model, and its roster now exceeds 650 participants.

CrowdStrike Holdings Inc. and Palo Alto Networks Inc. are among the organizations named for the program, which remains the only external route to Argon while Google continues safety work.

Koray Kavukcuoglu, Google’s chief AI architect, described the release as a phased process for high-capability systems.

The company is also using the U.S. voluntary channel for pre-release model access, so the cyber rollout is being handled as a staged deployment rather than a standard product availability announcement.

Argon’s performance evidence centers on long software tasks and business automation.

The model reached 77.9% on DeepSWE v1.1, a benchmark for extended software engineering work.

Google’s chart put Anthropic’s Claude Opus 5.5 at 74.2%, slightly above OpenAI’s GPT-6 Astra.

On AutomationBench, Argon led Opus 5.5 by 51.3% to 42.5%, and VentureBeat counted first-place results for Argon on 12 of the 18 benchmark charts.

The model can also keep far larger work products in a single run.

Its output ceiling is 1 million tokens; earlier Gemini systems stopped at 64,000.

That capacity is aimed at the same long-running engineering jobs represented by DeepSWE, and Google’s own teams have already put Argon agents into production-style internal work.

One internal effort uses the agents to convert C and C++ code into Rust.

The projects range from core libraries to Fuchsia’s Zircon kernel, which contains more than 800,000 lines.

In libgav1, Google’s open-source video decoder, agents took an existing Rust port and rebuilt 32,000 performance-critical lines as safe Rust that the compiler could optimize.

The revised decoder kept the same video output and ran 2.7 times faster than the earlier Rust version.

Argon agents are also being applied to memory efficiency.

Google engineers used fleetwide profiling data to identify waste, and the work has recovered over 300 tebibytes of capacity across company data centres.

The Fairwind edition removes cyber restrictions for approved defenders and internal Google teams.

In that mode, Argon can identify and remediate software flaws without step-by-step human handling.

It reached 68% on CWE-bench v1, tying GPT-6 Astra on the remediation test.

Google pairs that access with monitors for model reasoning traces and actions, designed to interrupt activity that departs from the user’s intended job.

Wiz Inc, Google’s cloud security unit, has already used Argon in Scan for Good, its free program for critical public systems.

In that work, the model identified a severe exposure in hospital software that revealed personal data, a flaw Google said previous frontier models failed to catch.

The wider commercial queue remains unresolved.

Paying developers and Google AI Ultra users are next, but no launch date has been set.

Introductory pricing is $2 for each million input tokens and $10 for each million output tokens, with cached input discounted by 95%.

After that period, Argon rises to $4 and $20, the same input and output rates Anthropic uses for Opus 5.5.