Google unveils Gemini 4 Argon: cyber defenders get it first, everyone else still has to wait

What Google announced

In a post on September 30, 2026, Koray Kavukcuoglu, Senior Vice President at Google DeepMind, announced Gemini 4 Argon. Google calls it a new "frontier" model, meaning its most advanced AI model. Google says Argon is built for long, multi-step work. Its focus areas are real-world software engineering, office knowledge work such as law and finance, and cybersecurity defense.

Google says that, for now, the model is available only to a group of trusted cyber defenders through the Fairwind Program. It says it is releasing Argon in stages and is taking part in the US government's voluntary process for giving access to models before release. Google says it will gather feedback from early testers and improve its safeguards before opening Argon to developers, businesses and consumers. Paid API customers will come first, meaning developers who pay to build the model into their own software, along with Google AI Ultra subscribers. The announcement gives no availability date.

Read the full description

Sources are collected, independently checked, then reviewed by Jev.

Capabilities and results as reported by Google

AI models are often compared using benchmarks, which are standardized tests that give a model a score. Google reports the following results for Argon:

• DeepSWE v1.1 (long-horizon software engineering): 77.9%, which Google calls a new state-of-the-art result.
• AutomationBench (Zapier's benchmark for carrying out business tasks from start to finish): 51.3%, which Google says ranks first. Google also says Argon leads the Vals Index.
• LVBench (understanding long videos): 91.7%.
• CWE-bench v1 (fixing security vulnerabilities): 68%, tied for first.

Google also describes how it uses Argon internally. It says thousands of employees use the model. Teams of Argon "agents" (AI that carries out tasks on its own) found memory savings in Google's data centers. Google says these freed more than 300 TiB of memory after rollout, with total savings estimated at 500 TiB to 1 PiB. Argon is also moving code written in C/C++ to Rust, a programming language designed to avoid many memory-related bugs. The largest of these projects is more than 800,000 lines in the Zircon kernel of Fuchsia OS, and the rewrites go through automated and human review before deployment. For libgav1, Google's open-source video decoding software, Google says the resulting memory-safe decoder runs 2.7 times faster than the existing Rust version ("port").

Cybersecurity: why defenders come first

Google says Argon can find, confirm and fix serious software security flags (vulnerabilities) on its own. It says trusted defenders and Google's own internal teams will get a version without the usual cybersecurity restrictions. As an example, Google says the security company Wiz is using Argon through Scan for Good, a Wiz program that protects critical public infrastructure for free. Through that program, Argon found a critical flaw that exposed sensitive personal data in medical software used by hospitals worldwide.

According to Google DeepMind, the Fairwind Program currently works with more than 650 partners worldwide. It gives priority to governments, operators of critical infrastructure and core technology platforms. Partners may use the model only for defensive and academic research. They must use phishing-resistant multi-factor authentication, which is a login that needs more than a password. They may give access only to internal security teams, and they may not resell it. Google DeepMind also says it runs background checks on organizations that apply.

Safety measures and related frameworks

Google says that before a broad rollout it will strengthen safeguards in four areas: 1. Preventing misuse, including cyber attacks and chemical, biological, radiological and nuclear attacks. 2. Defending against prompt injection, where hidden malicious instructions are used to hijack what a model does. 3. Watching for "misalignment", where a model pursues a task in ways that go beyond what the user intended. Google says it will monitor the model's step-by-step reasoning (its "chain of thought") and its actions, and stop it when necessary. 4. Hardening the isolated test environments ("sandboxes") in which the model is trained and evaluated. Google calls Argon its most resistant model to date against indirect prompt injection.

These measures build on frameworks Google DeepMind published earlier. The third version of its Frontier Safety Framework was released on September 22, 2025, and Tracked Capability Levels were added to it on April 17, 2026. The AI Control Roadmap, published on June 18, 2026, uses trusted AI systems as "supervisors" to review AI agents' reasoning, actions and plans.

What it means for general readers

• For now, ordinary users cannot use Argon. Google says paid API customers and Google AI Ultra subscribers will get access first, but it has not set a date.
• If Argon's flaw-fixing abilities work as Google claims, critical services such as hospitals and telecoms may fix security flaws sooner. That would indirectly protect the people who rely on those services.
• By releasing Argon to defenders first and the public later, Google signals that it treats powerful cyber capabilities as a technology that needs controls.
• For organizations that do not qualify for Fairwind, Google DeepMind says they can still use CodeMender, its code security agent, with publicly available models.

Browse the latest news in this topic

• Tech news