Google Launches Gemini 4 Argon Frontier Model

Google released Gemini 4 Argon on September 30, 2026, positioning it as the company’s next frontier model for software engineering, enterprise research and cybersecurity defense. SVP Koray Kavukcuoglu framed the model around sustained reasoning through long, complex workflows rather than quick single-turn answers. The company is not opening it widely yet. It first reaches a small group of trusted cyber defenders through what Google calls its Fairwind Program, with paid API customers and Google AI Ultra subscribers next in line once testing finishes.

Google says Argon already supports internal teams, pointing to work on quantum algorithm optimization, large scale migrations of C and C++ code into Rust, and memory efficiency work across its data centers that it says freed hundreds of terabytes of capacity. The company also widened Argon’s output limit sharply, moving from 64,000 tokens to 1 million tokens, a change meant to let the model work through long multi step problems in a single pass rather than breaking a task into smaller calls.

On cybersecurity, Google says Argon can find, check and patch vulnerabilities on its own, and it is releasing the model to internal teams and outside partners such as Wiz without the usual cyber safety limits, so they can draw on its full defensive capability. Google frames this as a response to a harder era of cyberattacks, and says the model found a flaw in hospital software that earlier models had missed.

Before opening Argon more broadly, Google lists four safeguard areas it is strengthening: blocking misuse for cyber or chemical, biological, radiological and nuclear purposes, resisting prompt injection attacks, watching the model’s own chain of thought for signs it is acting outside a user’s intent, and sealing the sandboxes used for training and testing. The company says it is also taking part in the US government’s voluntary process for pre-release model access while it works through these checks.

The numbers

• Input tokens carry an introductory price of $2 per million, rising to $4 per million once that period ends.
• Output tokens are priced at $10 per million during the same introductory window, climbing to $20 per million afterward.
• Argon scored 77.9% on DeepSWE v1.1, a benchmark of long horizon software engineering tasks, which Google calls a new state of the art result.
• On CWE-bench v1, which checks a model’s skill at fixing security flaws, Argon tied for first place with a score of 68%.

The staged rollout, starting with cyber defenders rather than the public, suggests Google wants outside pressure testing on a model it is marketing through professional and defensive use cases first, well before it reaches a consumer chat window.