Google starts Gemini 4 Argon rollout with selected cyber defenders

Google has begun rolling out Gemini 4 Argon to selected cyber defenders, giving a limited external group access to its new artificial intelligence model before any wider developer, enterprise or consumer release.

The company announced Argon on 30 September 2026 and is distributing it through Fairwind, its restricted cyber defence programme for governments and trusted partners. Reuters, in reporting published by Yahoo, also described the initial availability as limited to selected cybersecurity partners.

That makes Argon’s debut a controlled early deployment rather than a conventional public launch. Most developers and consumers cannot use the model, but it is too broad to say that nobody has access. Google has not identified the Argon testers, detailed how it selected them or provided a date for general availability.

Google says wider access will begin with paid API customers and Google AI Ultra subscribers. The company has not disclosed when that phase will start, where it will be available or when other enterprise and consumer users might follow.

Future pricing is published, despite no public API

Google has published prices intended to apply when Argon becomes more broadly available. Its planned introductory rates are $2 per million input tokens and $10 per million output tokens, with a 95 per cent discount for cached inputs. Google says the later rates will rise to $4 and $20 respectively.

Those figures do not mean a generally accessible paid API is operating now. They establish only that Google has set out future pricing alongside its intended rollout sequence.

The current official announcement therefore conflicts with an Ars Technica report stating that API pricing had not been announced. The timing of any update to Google’s page is unknown, so the available evidence does not establish whether those prices were visible when Ars published its account.

Google also documents a one million token output limit for Argon, compared with a previous limit of 64,000 output tokens. That could accommodate substantially longer generated results, but it should not be mistaken for a one million token input context window. Nor does output capacity by itself demonstrate accuracy, reliability or performance in long-running work. Google has not disclosed practical latency or quota limits.

Cyber performance evidence comes mainly from Google

Google positions Argon as a model for coding, enterprise work and cyber defence. Its published evaluation material reports a score of 77.9 per cent on DeepSWE v1.1 and 68 per cent on CWE-bench v1, where it says Argon tied for first place.

The company also says the model can autonomously find, validate and patch critical software vulnerabilities. These are substantial capability claims, but the reviewed public evidence does not include independent reproduction of Argon’s comparative performance. The benchmark results and safety assessments remain principally Google-reported findings.

Google has presented internal deployments as further evidence of practical value. It credits Argon-assisted work with freeing more than 300 TiB of data-centre memory and says the model is supporting projects to migrate more than 800,000 lines of the Fuchsia Zircon kernel from C and C++ to Rust. Those production outcomes have not been independently verified in the supplied material.

The most consequential security example is also the least transparent. Google, citing Wiz, says Argon discovered a critical vulnerability in healthcare software used by hospitals worldwide. It says the flaw exposed sensitive personal information and had been missed by previous frontier models.

Google’s announcement does not name the software or its vendor. It provides no CVE, affected versions, technical description, patch information or disclosure timeline. It also supplies no evidence of active exploitation or confirmation that anybody accessed data through the flaw.

The claim therefore cannot yet be assessed like a conventional public vulnerability disclosure. It should not be described as a breach, and the absence of product and remediation details prevents potentially affected organisations from determining whether the claim applies to their systems.

Reduced cyber restrictions, but not uncontrolled access

Google says selected defenders and its internal teams will receive Argon with its cyber guardrails removed. That does not establish unrestricted public access or an absence of other controls.

Fairwind, launched on 2 September 2026, is a limited-access programme that Google says has more than 650 participating partners. Its documented organisational requirements include restricting access to approved staff and using protections such as multi-factor authentication.

Google separately outlines planned safeguards, including refusals of harmful requests, defences against prompt injection, monitoring of model reasoning and actions, and hardened sandbox environments. It says monitoring can halt activity that moves beyond intended bounds, although the effectiveness of that system has not been independently established in the reviewed evidence.

The precise cyber restrictions removed for Argon partners remain undisclosed, as do the technical enforcement arrangements applied to individual participants. This distinction matters: selected defenders may receive capabilities unavailable to ordinary users while still operating within identity, access, monitoring and containment requirements.

Google describes Fairwind as a global programme aimed at governments, critical infrastructure and core technology platforms. That makes its access model relevant to Australian public-sector and critical-infrastructure organisations, but the reviewed material identifies no Australian partner, local rollout date or Australia-specific deployment.

For now, Argon is neither a generally available product nor a model that nobody can use. It is an early, restricted deployment with published future prices and extensive vendor claims. Google has yet to provide a wider release date, independent testing of the reported performance or enough detail to identify the healthcare software and its remediation status.

A restricted model rollout brings stated access controls, monitoring and containment into view, but leaves practical effectiveness as a separate question. With The Hack Academy's online cybersecurity training programme, you can study cybersecurity fundamentals and networking at your own pace, combining theory with practical exercises.