Google launches Gemini 4 Argon but restricts access to cybersecurity defenders first

Google on Thursday released Gemini 4 Argon, its most capable AI model to date, then immediately locked it behind one of the most restrictive release policies the industry has seen. General availability is not on the calendar. Instead, access goes first to roughly 650 partner organizations in Google's Fairwind Program — government agencies, critical infrastructure operators, and enterprise security teams — with paid API and Google AI Ultra subscribers next in line.

The rationale is straightforward. Argon was trained specifically to find, validate, and patch software vulnerabilities, and it is very good at it. Google says the model scored 68% on CWE-bench v1, a vulnerability remediation benchmark, and 77.9% on DeepSWE v1.1 for extended software engineering tasks. In one documented case, Argon identified a critical flaw in healthcare software used by hospitals after earlier frontier models had failed to detect it.

The concern is the same capability that makes Argon useful to defenders makes it dangerous in other hands. The phased rollout is Google's answer to that tension: trusted cybersecurity professionals get guardrail-free access for defensive work, while the broader public waits for additional safety work to complete.

Argon also significantly expands on its predecessor. Maximum output jumps from 64,000 to one million tokens, enabling the model to reason across entire codebases in a single pass. Google says it also hardened the model against indirect prompt injection and ran isolated red-team exercises targeting chemical, biological, radiological, and nuclear risk scenarios before release.

The model replaces the previously announced Gemini 3.5 Pro, which Google cancelled earlier this year. Benchmark comparisons against OpenAI's Astra and Anthropic's Opus show Argon competitive on coding and cyber tasks, though Google acknowledges performance varies by benchmark. A broader release timeline has not been disclosed.

The Fairwind Program launched September 2, 2026, and requires participants to restrict Argon access to employees in cybersecurity, incident response, or penetration testing roles, with multi-factor authentication as a baseline control. As reported by TechWire Asia, the program had over 650 participating partners globally at the time of Argon's launch.