Google unveils Gemini 4 Argon: Its most powerful AI model yet, focused on coding and cyber defence

Google unveiled Gemini 4 Argon, its latest frontier AI model, on Wednesday, 30 September. The model is intended for more complex tasks such as software development, investigation, enterprise work and writing. Google claims Argon's greatest strength is in defensive cybersecurity, where it can automatically discover, approve and fix software vulnerabilities.

The model is not yet available to the public. Google's new Gemini 4 Argon is initially being released to a select group of its trusted cyber defenders through its Fairwind Program. The phased release will enable Google to enhance its safety measures before bringing Argon to developers, enterprises and consumers, Google said. The company said paid API users and Google AI Ultra users will be the first to see wider adoption.

Gemini 4 Argon targets complex coding tasks

Google says Argon is already being used internally by thousands of its development team members in debugging, research and large-scale software projects. The company said the model has helped them optimize quantum computing algorithms and identify memory improvements in their data centres.

Argon is also being used for major code migrations. Google claims to be helping shift C and C++ code bases to Rust, even for projects with hundreds of thousands of lines of code.

Google said Argon replaced 32,000 lines of SIMD code in the libgav1 video decoder and created a Rust port that is 2.7 times faster than a previous port but supports the same video output.

1 million-token context for longer workflows

Google has increased Argon's output to one million tokens compared to the previous-generation Argon, which had an output of 64,000 tokens. The company says this allows the model to sustain longer reasoning processes and tackle complicated tasks in a single workflow.

Google claims Argon achieved 77.9% on the DeepSWE v1.1 benchmark that assesses the performance of software engineering tasks with long time horizons. It also has a 51.3% rating from Zapier's AutomationBench. It's a model that can also process visual input. Google claims it can process professional charts, long videos and many documents.

Google positions Argon as a cybersecurity tool

One of the key priorities of the new model is cybersecurity. According to Google, Argon can find and fix important bugs with minimal human participation.

On CWE-bench v1, Argon had a score of 68% and tied for first place. Cybersecurity firm Wiz is also running the test as part of its Scan for Good programme. Google claims Argon discovered an important flaw in healthcare software that the previous frontier models overlooked. Google will launch Argon without the usual cyber guardrails, allowing trusted cyber defenders to conduct extensive testing.

Safety checks before wider release

Google said it is strengthening safeguards against cyber misuse, CBRN-related misuse and prompt injection attacks. The company is also installing systems to continuously track model behaviour and to take action when required.

According to Google, Gemini 4 Argon will start at $2 per million input tokens and $10 per million output tokens, with a 95% discount for input tokens that have been cached.

The company has not announced a final date for full consumer availability.