Google Unveils Gemini 4 Argon, Giving Cyber Defenders First Access

Google unveiled a new AI model, Gemini 4 Argon, on September 30 local time. The company is positioning it for complex software development, legal and financial work, and cybersecurity defense. For now, it is being rolled out first, in stages, to cyber defenders taking part in the Fairwind Program.

What changes on long tasks

The change Google emphasizes most is the model’s ability to carry multi-step, long-running tasks through to completion. The company says it has raised the length of a single response from 64,000 tokens to 1 million tokens. Tokens are the fragments of text an AI uses to process language. The upshot, according to Google, is that the model now has room to think longer and write out far longer results within a single task.

For development work, Google says the model can handle everything from bug fixes to large-scale program overhauls. The company reports a score of 77.9% on DeepSWE v1.1, a benchmark that measures performance on long development tasks. It also says Argon took first place on Zapier’s AutomationBench, which tests whether a model can carry core business workflows from start to finish, with a score of 51.3%.

Google also shared examples of using the model internally. The company says it used Argon to improve the existing Rust version of libgav1, video decoding software, raising processing speed by a factor of 2.7. The video output remained identical. In its data centers, Google says an agent - an AI that handles work on its own - tuned memory usage, and that applying the change would free up at least 300TiB of memory. When code changes are substantial, they go through automated checks and human review before reaching production.

Why cyber defense comes first

Google says that releasing powerful capabilities safely requires a staged approach. The company claims Argon can find serious software vulnerabilities on its own, verify whether they are real problems, and then fix them. For trusted defenders and its own internal teams, Google plans to provide a version without cyber-related safety restrictions.

Source: Google (press release)

The security firm Wiz is using Argon in Scan for Good, an initiative that protects public infrastructure free of charge. According to Google, Argon identified a vulnerability in medical software used by hospitals around the world that could have leaked sensitive personal information. Google says earlier models missed the issue. Before widening access, the company is also strengthening measures against misuse, along with defenses that filter out malicious instructions hidden inside external documents. It is adding monitoring that halts execution if the model strays from the user’s intent.

When, and at what price

Google plans to widen access to developers, businesses, and general consumers, starting with paying API customers and Google AI Ultra subscribers. An API is the mechanism developers use to plug a model into their own services. The company says it will expand the rollout while refining safeguards based on feedback from early testers.

Source: Google (press release)

Launch pricing is $2 per million input tokens and $10 per million output tokens. Reusing cached input cuts the input price by 95%. For services that repeatedly feed the model the same material, that adds up to a large difference in cost.

Google wants to extend Argon’s reach across long-running tasks in development, office work, and security. The open questions are whether the capabilities highlighted in internal examples and benchmarks hold up in real users’ work, and how far the company will widen its rollout as it refines its safeguards.