Gemini 4 Argon: Google's new AI model for cybersecurity and code

What's new

On September 30, 2026, Google unveiled Gemini 4 Argon, the company's newest and most advanced AI model. Argon is built to handle long, multistep tasks that require sustained focus, such as software development, financial analysis and IT security. One of the big pieces of news is that the model can work with up to 1 million tokens in its output, meaning the text it writes itself, compared with only 64,000 in previous models. Tokens are the small pieces of text AI models compute in, and more tokens mean Argon can keep track of much longer and more complex tasks without losing the thread.

What's clever

The cleverest thing about Argon is that it is trained specifically for cybersecurity. It can find security holes in software on its own, confirm that they are real, and fix them, a kind of automatic digital watchdog for companies' IT systems. Google has also made the model better at resisting so-called prompt injection, where malicious messages try to trick the AI into doing something unwanted. Argon is also strong at understanding images, videos and documents, which makes it useful for everything from long reports to video analysis.

Cheaper or better?

On benchmarks, meaning standardized tests, Argon beats its competitors on most of the 18 measurements Google itself has published. On the software test DeepSWE, Argon scores 77.9 percent versus 74.2 percent for Anthropic's Claude Opus 5.5 and 74.1 percent for OpenAI's GPT-6 Astra. On a legal test called Harvey's Legal Agent, the lead is even bigger: 19.6 percent versus under 6 percent for the two competitors. On the other hand, Argon loses on some tests, including to GPT-6 Astra on another coding test.

The price is set at $2 per million input tokens and $10 per million output tokens for an introductory period, which is about a fifth of what GPT-6 Astra costs. Later, the price rises to $4 and $20, matching Claude Opus 5.5. These are prices for developers using the model via Google's API, not a subscription price for individual users.

What it's good at

Argon is especially good at long, multistep tasks: writing and fixing large amounts of code, analyzing financial data, understanding long videos and documents, and not least finding and closing security holes before hackers get to them. Right now, the model is only available to selected cybersecurity partners in Google's so-called Fairwind program, which counts more than 650 companies. Google has not yet set a date for when regular developers, businesses or Google AI Ultra subscribers will get access, but the company says it will happen "as quickly as possible."