Google gives Gemini 4 Argon to cyber defenders before public release

Why it matters

Google is pairing claims of long-running coding and vulnerability-remediation ability with restricted early access, making the safeguards and eventual public availability part of the product's commercial test.

Google is giving its new Gemini 4 Argon model to selected cyber defenders before opening access to developers, enterprises and consumers, CEO Sundar Pichai (@sundarpichai) said on September 30th. In his X announcement, Pichai said Google teams were already using Argon for work ranging from coding to quantum computing. The company has not set a date for broader access.

The rollout puts cybersecurity at the front of the launch. Google says Argon can find, validate and patch software vulnerabilities, and that the model will go to approved users through its Fairwind Program. Google describes Fairwind as a controlled-access program for trusted defenders, including governments and critical-infrastructure operators. Its stated rules restrict use to defensive and research work, require access controls and prohibit partners from sharing or reselling access.

In a Google DeepMind announcement published the same day, Google said it was also participating in the U.S. government's voluntary process for pre-release model access. Google says it will continue testing safeguards with early users before expanding availability. The announcement says broader access will start with paid API customers and Google AI Ultra subscribers, without a public timetable.

That caution sits beside a more consequential product claim: Google says it will give internal teams and Fairwind defenders access to Argon without cyber guardrails, so they can use its full cybersecurity capabilities. For approved defenders, that could mean letting the model work more freely on vulnerability discovery and remediation. It also makes vetting, monitoring and the program's permitted-use rules central to the launch, rather than details that can be left to a later enterprise policy.

Google says Argon can produce up to 1 million output tokens, compared with 64,000 for the previous limit. The company is pitching that capacity for extended, multi-step work rather than ordinary question-and-answer use. Its reported benchmark results include 77.9% on DeepSWE v1.1 for long-horizon software engineering, 51.3% on Zapier's AutomationBench, and 91.7% on LVBench for long-video understanding. Google also says Argon tied for first on CWE-bench v1, a vulnerability-remediation test, with a score of 68%.

Those are Google's reported results, not independent confirmation of a general performance lead. Benchmarks measure particular tasks under particular test conditions; they do not establish how reliably a model will perform across customer codebases or security operations. The company did not publish a public release date alongside the results.

Google's launch case also rests on internal use. The company says thousands of Googlers are using Argon. It reports that a team used model agents to identify memory optimizations that freed more than 300 TiB across its data centers after rollout, with estimated total savings of 500 TiB to 1 PiB. Those totals combine a reported deployed result with a larger estimate, and Google did not describe them as an independently audited measurement.

Another example is a rewrite of Google's open-source libgav1 video decoder. Google says Argon agents replaced 32,000 lines of SIMD code in an existing Rust port, producing a version that ran 2.7 times faster than that Rust port while returning identical video output. The comparison is with the prior Rust implementation, not a claim that it outperformed the optimized C++ implementation; Google says the result brought the Rust version closer to it.

The commercial terms are unusually concrete for a model that is not yet broadly available. Google says introductory API pricing will be $2 per million input tokens and $10 per million output tokens. After that introductory period, the announced prices rise to $4 and $20, respectively. Google has not said when the introductory period begins or how long it lasts.

Pichai's post called the release an early look, and the company has attached an explicit condition to the public rollout: access will expand as Google gathers feedback and strengthens its safeguards. For now, the launch combines a high-capability model, company-reported internal productivity examples and a security program designed to put early access in vetted hands. The next meaningful test will come when developers can run Argon on their own systems and compare its performance, cost and security behavior against their current tools.