Google has unveiled Gemini 4 Argon, its newest top-tier model, and handed it first to a small group of vetted security teams rather than to the public. In a post on Google’s official blog dated 30 September, Koray Kavukcuoglu said the model is rolling out through the company’s Fairwind Program to “trusted cyber defenders,” with developers, businesses and consumers to follow after what Google calls a phased approach.
That sequencing is the real news. Google describes Argon as strong at finding, confirming and patching serious software vulnerabilities, and says defenders and its own internal teams will get a version without the usual cyber guardrails. Holding back a model that can hunt for security holes is a different posture from the usual launch day, when everyone gets access at once.
Google says it is also enrolled in a voluntary U.S. government program that covers access to models before launch, and that it will widen availability gradually. It gives no date for the wider launch. When it comes, Google says paid API customers and Google AI Ultra subscribers will be first in line. The company has published launch pricing: $2 for every million tokens a customer sends in, $10 for every million the model sends back, and a 95 percent discount on cached input.
The performance claims are Google’s own, and the post leans on third-party benchmarks without showing independent verification. Google says Argon sets a new best score of 77.9 percent on DeepSWE v1.1, a test of long software engineering tasks, and tops the Vals Index, which scores models on tax, legal, coding and finance tasks and weights each field by its share of U.S. GDP. It also reports 51.3 percent on Zapier’s AutomationBench, 91.7 percent on the LVBench long-video test, and a 68 percent tie for first on CWE-bench v1, which measures how well a model fixes security bugs.
The internal examples are more concrete. Google says teams of Argon agents analyzed performance data from its data centers and found memory savings it puts at over 300 TiB once the changes ship, with total savings estimated at 500 TiB to 1 PiB. In a separate project, it says the model helped rewrite part of the libgav1 video decoder in memory-safe Rust, and the result, it says, runs 2.7 times faster than the previous Rust port while its output matches the old version exactly. Google also says Argon beat a published baseline by 40 percent when optimizing a quantum computing routine, in a matter of minutes. None of these results has been checked outside the company.
One outside name appears in the post. Google says the security firm Wiz is already using Argon in its Scan for Good program, which looks for exposures in critical public infrastructure at no charge, and that the model found a serious flaw, one exposing personal data, in healthcare software deployed in hospitals around the world, which earlier frontier models had missed. The post does not name the software or say whether the flaw has been fixed.
Argon also raises its maximum output to 1 million tokens, up from 64,000, which gives it room to work through a hard problem in a single pass. For agent builders, that matters more than another benchmark point: longer unbroken runs mean fewer hand-offs between model calls.
Google lists four safeguard areas it is still strengthening: blocking misuse for cyber and chemical, biological, radiological and nuclear attacks, resisting prompt injection (hidden instructions planted in content to hijack a model), monitoring the model’s reasoning and actions for misalignment, and locking down the sandboxes where risky tests run. It says it deliberately kept monitoring findings out of training so the model would not learn to dodge the monitors.
The strategic bet is visible in the structure. By pairing a staged release with a security-first audience, Google positions Argon as a tool for defense before anyone can point it at an attack, and it gives regulators a record of caution. The cost is time: rivals shipping broadly in the interim get the developers.
Security teams outside Fairwind have no way to test these claims yet. Until Google sets a release date and publishes independent results, the sensible move is to treat Argon’s numbers as marketing and plan around the price list.
Reported by Google’s official blog on 30 September 2026.
