Google begins restricted rollout of Gemini 4 Argon

Google announced Gemini 4 Argon on September 30, 2026, and began a restricted rollout to selected cybersecurity defenders through its Fairwind Program. Google said broader access would follow testing, beginning with paid API customers and Google AI Ultra subscribers. It gave no public-release date in its announcement.

Access for security work

Google says selected Fairwind partners can use Argon for vulnerability research and patching, including through CodeMender. The Fairwind Program lists more than 650 partners globally. Google has not said how many of those partners have received Argon access, or that all program members are eligible to use it now.

The rollout gives selected defenders an opportunity to test the model before general access. It also limits opportunities for developers and researchers outside the program to assess its behavior independently. Google has not said whether capabilities or safeguards will differ when access expands.

What the benchmarks show

On Collinear AI’s 120-task CWE-bench v1, Argon tied Grok 4.7 and GPT-6 Astra at 68% on the programmatic single-run measure. The models ran with different agent harnesses, which limits what their relative scores establish about the models alone. The test measures performance on its specified tasks; it does not measure whether patches are safely deployed in production.

Vals AI ranked Argon first on its September 30 Vals Index, with a score of 68.90%, ahead of Claude Sonnet 5.5 at 67.04%. Artificial Analysis scored Argon 53 on its Intelligence Index, tied with GPT-6 Astra. The independently published results support describing Argon as competitive on the tested tasks. They do not establish a general lead across tasks or a measurable gain in software-production efficiency.

Operational claims and remaining questions

Google says Argon can autonomously find, validate and patch critical vulnerabilities. It also says Wiz used the model to uncover a serious healthcare-software exposure. The announcement does not provide enough public detail to independently assess that case or its remediation. Google’s claims about internal code migrations, infrastructure savings and results on internal vulnerability tests likewise remain company-reported; the cited independent scores do not verify them.

Accounts of Argon’s coding performance inside Google also differ. Bloomberg reported that some employees found it uneven on real-world coding tasks. Google called the characterization of coding underperformance inaccurate, while another employee told Bloomberg there was broad internal agreement that Argon was frontier-level. The available benchmark results do not resolve how consistently it performs across production codebases.

Google announced introductory API prices of $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period. It did not specify when that period ends. Alongside the release date, the number of defenders with access and the outcomes of their early security work remain open questions for prospective users.