Google DeepMind has unveiled a new frontier AI model. The model is called Gemini 4 Argon. Google announced it on September 30, 2026, according to CNBC.
Argon targets long, complex workloads across several domains. Google built it for coding and enterprise knowledge work. It also targets cybersecurity defense tasks directly. The company says it can sustain reasoning across many steps.
Gemini 4 Argon Targets Cyber Defense and Long Coding Tasks
Argon follows Gemini 3.8 Live, the company’s previous flagship release. It is the first model in the Gemini 4 family. DeepMind positions it for legal work and finance research too.
Google says Argon can autonomously find software vulnerabilities. It can also validate and patch those flaws. That builds on earlier work like Gemini 3.8 Flash Cyber, per NeuralTrust.

Output Window Jumps From 64,000 to 1 Million Tokens
The headline technical change is output length. Earlier Gemini models capped out at 64,000 tokens. Argon raises that limit to 1 million tokens, per NeuralTrust.
That jump changes how complex work gets handled. A large code refactor no longer needs many prompts. One single response can carry the entire task. Google also made repeated context much cheaper to run.
Cached input tokens now get a 95 percent discount. That pricing move is detailed by NeuralTrust. It makes long, iterative sessions far more affordable.
Fairwind Program Limits Early Access to Vetted Defenders
Argon is not yet open to the public. Access starts with a narrow group of testers. They reach the model through Google’s Fairwind Program, per NeuralTrust.

Fairwind itself launched back in early September. It pairs Gemini models with Google’s CodeMender repair tooling, according to SecurityWeek. More than 650 partners already use that broader program.
Google is also moving carefully on policy matters. Argon is undergoing the U.S. government’s voluntary safety review. Wider access is expected after that review ends, per NeuralTrust and CNBC.
Why This Gradual Rollout Actually Matters
A one-million-token output window is a real jump. It is not a small footnote. It is a structural shift in model capability.
Google is pairing that power with real caution. Cyber-capable models carry genuine dual-use risk by nature. Gating access to vetted defenders reflects that tension, per SecurityWeek.
Watch for Google’s next move once reviews finish. Expanded access to developers may follow afterward. Pricing and safety guardrails will be worth checking first. For now, Argon stays limited to trusted Fairwind testers only.
