Google launches Gemini 4 Argon frontier model, restricting initial access to vetted cyber defenders

A deliberately narrow first release

Google has announced Gemini 4 Argon, its next frontier AI model, but instead of a broad launch the company is restricting initial access to a vetted group of cyber defenders, with general availability to follow after further safety work.

According to Google's announcement, Argon is rolling out through the company's Fairwind Program to a set of trusted cyber defenders. Google says releasing frontier capabilities at this level calls for a phased approach, and that it is taking part in the U.S. government's voluntary process for pre-release model access while it gradually widens availability. The Verge attributes those remarks to Koray Kavukcuoglu, Google's chief AI architect and DeepMind SVP, who took over leadership of DeepMind in August.

Google says it will collect feedback from early testers and iterate on guardrails before opening Argon up to developers, enterprises and consumers. The Verge reports that the safeguards under construction include defenses against misuse and prompt injection attacks, along with monitoring for signs of misalignment.

What Google claims the model can do

Google positions Argon as a model built for long-horizon reasoning: sustaining deep, multi-step work rather than producing one-shot answers. The output token limit rises to 1 million tokens, up from 64K in the previous generation, which Google says gives the model room to work through hard problems within a single run.

On benchmarks, Google reports state-of-the-art results: 77.9% on DeepSWE v1.1, which measures real-world long-horizon software engineering; the leading position on the Vals Index, which weights finance, coding, legal and tax performance by contribution to U.S. GDP; first place on Zapier's AutomationBench at 51.3%; and 91.7% on LVBench for long-video understanding. The Verge notes that Google also published comparison charts placing Gemini 4 ahead of competing models from OpenAI and Anthropic, after what appeared to be leaked benchmark results circulated on X earlier in the week.

Google says the model is already in internal use. Claimed examples include beating a published quantum-computing baseline by 40% during subroutine optimization, freeing over 300 TiB of data-center memory through agent-driven optimizations with an estimated 500 TiB to 1 PiB in total savings, and C/C++-to-Rust migrations at scale — up to more than 800,000 lines for the Fuchsia OS Zircon kernel, with a rewritten libgav1 video decoder running 2.7x faster than the previous Rust port.

When Argon does become broadly available, it will carry an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%.

The cybersecurity angle

The most unusual element of the launch is cyber defense. Google says Argon can autonomously find, validate and patch critical software vulnerabilities, and that for trusted defenders and Google's own internal teams it is being released without cyber guardrails so those full capabilities can be used.

Wiz is already deploying Argon through its Scan for Good initiative, which finds and remediates high-risk exposures in critical public infrastructure at no charge. In an early demonstration, Google says the model uncovered a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals worldwide — a severe risk that earlier frontier models had missed.

Why it matters

Two things make this launch a signal beyond the usual benchmark race. First, the restriction itself: a frontier model being withheld from general availability because its defensive — and implicitly offensive — cyber capabilities are considered too risky to ship openly sets a notable precedent for how labs handle dual-use AI. Second, Google's participation in a voluntary U.S. government pre-release review process, however informal, sketches the outline of a release norm taking shape between major labs and the state.

The competitive timing sharpens the point. As The Verge observes, Argon arrives just one day after OpenAI's DevDay, where OpenAI introduced its Dots AI agent and GPT-6.1 Sol model — and separately confirmed it would not ship a planned GPT-6.1 Astra model over safety concerns. The two leading labs are now visibly treating capability and restraint as parts of the same launch story, and how quickly Google expands Argon access will be a test of whether that restraint holds.