Google’s Gemini 4 Argon Starts With Cyber Defenders, Not the Public

Google announced Gemini 4 Argon on September 30, 2026, but most developers and Gemini subscribers cannot use the new frontier model yet. Its initial rollout is limited to selected cyber defenders and trusted testers participating in Google’s Fairwind Program.

The restricted release is central to the product, not a minor availability detail. Google says Argon can autonomously find, validate and patch critical software vulnerabilities. For approved defenders and Google’s internal security teams, the company is offering a configuration without its normal cyber guardrails, allowing authorized specialists to use the model’s full security capabilities.

Google plans to expand Argon access after collecting feedback and strengthening safeguards. Paid Gemini API customers and Google AI Ultra subscribers are named as the first groups in line, followed by broader developer, enterprise and consumer availability. The company has not announced a date for any of those stages.

As of October 1, Google’s public Gemini API pricing documentation does not list an Argon endpoint. That means the prices announced alongside the model describe its planned commercial terms, not an API that ordinary developers can call today.

Gemini 4 Argon availability and pricing

Google has not said how long the introductory pricing will last, which regions will receive API access first or what rate limits and product integrations will accompany the model. It has also not confirmed whether Ultra subscribers will receive Argon in the main Gemini app, a coding product or a separate interface.

The one-million-token output ceiling is unusually large. At the introductory rate, a hypothetical response using the full allowance would incur $10 in output-token charges before input, tools or other infrastructure costs. The comparable charge would rise to $20 after the introductory period.

That limit should not be confused with a guarantee that million-token responses will be practical or reliable. Google describes it as headroom for long reasoning and generation trajectories. Independent testing will be needed to measure completion rates, latency, consistency and cost on workflows that actually approach the ceiling.

▶Play video

Why Fairwind partners get Argon first

Google launched Fairwind on September 2 as a controlled-access program for advanced cyber-defense systems. It initially paired the security-focused Gemini 3.8 Flash Cyber model with CodeMender, Google’s agent for finding and repairing software vulnerabilities. Argon now becomes the program’s most capable model.

Fairwind works with more than 650 partners globally, but Google does not say all of them are receiving Argon. Its announcement refers to a set or initial cohort of participating defenders, indicating that the model’s early distribution is narrower than Fairwind’s total membership.

The program prioritizes three kinds of organizations:

• Governments and national cyber authorities responsible for public networks and citizen services.
• Critical infrastructure operators in areas such as healthcare, telecommunications, energy and finance.
• Core technology platforms whose software and services affect large numbers of downstream users.

Academic laboratories focused on defensive benchmarking may also apply. Google says it performs background checks covering applicants’ security histories and records of ethical operation.

Approved organizations must restrict model access to internal cybersecurity, incident-response or penetration-testing teams. Fairwind’s terms also call for individual authentication, phishing-resistant multifactor authentication, appropriate access controls and employee-usage tracking. Partners cannot resell, redistribute or share their access.

Authorized threat simulation, reverse engineering and malware analysis are permitted when conducted for defensive or academic research. Malicious malware creation is prohibited. Google also says Fairwind customers can obtain zero-data-retention support when Argon is accessed as a managed model through Gemini Enterprise.

Trusted defenders receive broader cyber capabilities

The most consequential distinction is Google’s decision to provide approved defenders with Argon “without cyber guardrails.” The goal is to let qualified teams use the model’s complete vulnerability research and exploitation knowledge when examining systems they are authorized to protect.

Outside that controlled setting, Google says Argon is designed to refuse harmful requests involving cyberattacks or chemical, biological, radiological and nuclear threats. Its wider release is therefore waiting on additional work in four areas: misuse prevention, resistance to indirect prompt injection, monitoring for behavior that goes beyond a user’s intentions and stronger isolation for high-risk testing environments.

Fairwind shifts part of the safety burden from model refusals to organizational controls. Instead of relying exclusively on the model to decide which security requests are legitimate, Google vets the customer, limits which employees can use the system and imposes contractual and technical restrictions.

The approach gives defenders an opportunity to find and repair vulnerabilities before similar capabilities become broadly available. It also creates a transparency challenge: researchers and ordinary customers cannot independently inspect the most capable cyber configuration while it remains limited to approved organizations.

What Google says Gemini 4 Argon can do

Argon is positioned as a general frontier model rather than a narrowly specialized security tool. Google says it is designed for long-running software engineering, financial research, legal work, visual analysis, writing and other professional workflows.

The company reports a score of 77.9% on DeepSWE v1.1, which evaluates long-horizon software engineering. Google also reports 51.3% on Zapier’s AutomationBench, 91.7% on the long-video-understanding benchmark LVBench and a first-place tie at 68% on CWE-bench v1, an evaluation of vulnerability remediation.

Independent evaluator Vals ranked Argon first on its overall index at launch, with a score of 68.9%. The same evaluation showed a more mixed picture across individual tasks: Argon ranked highly in finance, code migration and several professional-work tests, but placed lower on some computer-use, terminal and medical-documentation evaluations.

Those results support describing Argon as broadly competitive, not universally superior. They also do not replace public product testing. Google’s initial distribution means most developers cannot yet compare the model on existing codebases, unreliable tools, ambiguous requirements or other conditions that make production work harder than benchmark tasks.

Axios reported that Bloomberg had spoken with Google employees who questioned how well the model handled some real-world coding work despite its benchmark scores. Google disputed that characterization. The disagreement is another reason to treat launch-day rankings as provisional until more independent users can test the model.

▶Play video

Argon is already being used inside Google

Google says thousands of employees have tested Argon for coding, research and writing. Its announcement highlights several projects intended to demonstrate the model’s ability to sustain long engineering workflows.

• Argon agents analyzed fleet-wide profiling data and identified memory optimizations that Google says freed more than 300 TiB across its data centers, with estimated potential savings of 500 TiB to one PiB.
• The model has assisted migrations from C and C++ to Rust, including work involving more than 800,000 lines associated with the Fuchsia operating system’s Zircon kernel.
• For Google’s libgav1 video decoder, Argon reportedly replaced 32,000 lines of specialized SIMD code in an existing Rust port. Google says the result ran 2.7 times faster than that port while producing identical video output.
• Google says the model improved the resource efficiency of a quantum-computing algorithm by 40% compared with a published baseline.

These remain company-reported case studies, and Google says safety-critical migrations are undergoing automated testing, emulation, manual auditing and review before production use. The examples present Argon as an engineering accelerator rather than a system authorized to approve its own production changes.

Google also says security company Wiz used Argon through its Scan for Good initiative to uncover a critical vulnerability in healthcare software used by hospitals internationally. The company says earlier frontier models missed the flaw, but neither Google nor Wiz has publicly identified the affected software or released enough technical information for outsiders to reproduce the result.

What happens next

The next major milestone will be an actual availability date. Google still needs to specify when API customers and Ultra subscribers will receive access, which countries will be supported, how Argon will appear in consumer products and whether its most powerful cyber features will remain exclusive to Fairwind.

Developers will also need details about rate limits, tool support, latency, data policies and the duration of introductory pricing. Independent testing should establish whether the one-million-token output allowance improves the completion of long tasks enough to justify the potential cost.

For now, Gemini 4 Argon is a significant model announcement but an intentionally limited product release. Google has disclosed its planned pricing, several performance claims and its first customer groups, while keeping the model’s most sensitive configuration inside a vetted cyber-defense network.