Google says Gemini 4 Argon is rolling out to selected cybersecurity defenders. Broader access for developers, enterprises, and consumers is still pending, so readers looking for a new Gemini model to try should treat this as a restricted release.
The Gemini 4 Argon announcement, dated September 30, 2026, describes a model built for extended software engineering, professional knowledge work, and vulnerability research. Early distribution is through Google DeepMind’s Fairwind Program.
Wiz separately reports using Argon in its Scan for Good initiative. That provides a partner account of deployment without independently establishing Google’s performance claims. The available evidence supports controlled access and reported capabilities; it does not establish broad availability or independently reproduced results.
Google Describes Early Access, Not a Public Launch
Argon is reaching “a set of trusted cyber defenders,” according to Google’s announcement. The company promises to expand availability to developers, enterprises, and consumers after further feedback and work on safeguards, but gives no firm public-release date.
Selected organizations receiving a model does not mean ordinary Gemini users can access it or developers can immediately select it through a generally available service.
September 30 is the stated event date available in the evidence. No publication clock time is supplied, so there is no basis for assigning an exact UTC announcement time. Wiz’s update carries the same calendar date, which leaves the order of the two statements unclear.
Google also lists planned introductory token pricing:
- Input: $2 per million tokens.
- Output: $10 per million tokens.
- Cached input: 95% below the input rate, equivalent to $0.10 per million tokens.
The figures describe a stated launch offer. An open purchasing route has yet to be established, and it remains unclear when people outside the restricted rollout will be able to use the model at those rates.
Fairwind Membership Does Not Guarantee Argon Access
The Fairwind Program aims to give high-priority defenders early access to advanced models before comparable capabilities become widely accessible. Its priority groups include governments and national cyber authorities, critical infrastructure operators, and core technology platforms. Healthcare, telecommunications, energy, and financial networks are among the sectors identified.
Fairwind says it works with more than 650 partners globally, but specifies that a set of partners receives exclusive Argon access. That wording establishes neither access for every partner nor automatic access upon joining.
Applicants face organizational due diligence, including checks on their security history and record of ethical operations. There is no unrestricted signup.
Approved organizations must also manage access internally. The program requires user-level authentication, phishing-resistant multifactor authentication, and applicable access controls. Organizations may grant Argon access only to their internal cybersecurity, incident-response, or penetration-testing teams, and must track employee access and use.
Partners cannot share, redistribute, or sell access. Academic labs focused on defensive benchmarking are welcome to apply, though eligibility does not guarantee approval. Fairwind is an organizational access program, with no subscription described that any individual researcher can purchase.
Argon’s Cybersecurity Pitch Combines Discovery and Patching
Google says Argon can autonomously find, validate, and patch critical software vulnerabilities. Fairwind partners can use it as a standalone model or through CodeMender, which the program describes as a specialized agent for vulnerability research and software fixes.
A usable security workflow needs more than model reasoning: authorized access to code or systems, a way to test proposed changes, and a process for reviewing findings before they affect production. CodeMender is presented as a way to provide that workflow without requiring defenders to build and maintain their own harnesses. This is an advertised benefit, with no assurance that every proposed patch will be correct or safe.
One consequential detail is Google’s statement that trusted defenders and its internal teams will receive Argon “without cyber guardrails.” Its stated rationale is to let approved users access the model’s full cybersecurity capabilities.
Authorization requirements remain in place. Fairwind allows dual-use work such as authorized threat simulation, reverse engineering, and malware analysis for defensive or academic purposes. It explicitly prohibits malicious tasks, including creating malware.
Removing certain model-level cyber restrictions for vetted users does not authorize them to test someone else’s infrastructure.
Wiz Reports Using Argon, but Earlier Results Need Careful Attribution
Wiz’s Scan for Good update, dated September 30, 2026, says the initiative now uses Gemini 4 Argon alongside Gemini 3.8 Flash Cyber. This corroborates a partner-reported deployment, but offers no independent benchmark of Argon and does not isolate its contribution from other models, tools, or human researchers involved.
Scan for Good focuses on public services, critical infrastructure, nonprofits, and other organizations where a security failure could have substantial consequences. Wiz describes combining AI with human security researchers to uncover public exposures and complex attack paths.
Examples include hospital systems, municipal services, transportation infrastructure, and technology platforms. The update does not establish which earlier findings were made using Argon, so the initiative’s reported remediation results should not be presented as Argon-specific achievements.
Wiz also sets out boundaries for its work:
Testing requires authorization through an applicable bug-bounty program, an established vulnerability-disclosure policy, or explicit permission. Potential findings receive human review, and affected organizations are contacted privately.
That workflow gives the deployment report more practical context than a headline vulnerability count would. Findings must be validated, disclosed responsibly, and translated into fixes without creating additional harm.
Google’s Benchmark Claims Leave Important Questions Open
Google reports a 77.9% score on DeepSWE v1.1, which it describes as measuring long-horizon software engineering, and a 68% score on CWE-bench v1, which evaluates vulnerability remediation. It says Argon ties for first place on the latter.
The supplied sources provide neither an independent evaluation nor enough methodological detail to establish how closely these Google-reported scores predict performance in a particular organization’s environment.
Frequently Asked Questions
3 questions
1Can individuals get access to Gemini 4 Argon?
The supplied sources do not describe general individual access. Early access is restricted to selected, vetted Fairwind organizations and their authorized teams. Academic labs focused on defensive benchmarking may apply, but eligibility does not guarantee approval. Google’s announcement gives no firm public-release date.
2
Sources
- Gemini 4 Argon announcementblog.google
- Fairwind Programdeepmind.google
- Scan for Good updatewiz.io
