Google's Gemini 4 Argon AI Model Released, But Access is Restricted

Google has released its new AI model, Gemini 4 Argon, which is designed for long-horizon work such as software engineering, legal and financial research, and cyber defense. The model's key feature is its ability to reason and write hundreds of thousands of tokens in a single pass.

The output ceiling of the model has risen to 1 million tokens, up from 64,000, making it more efficient for tasks such as porting C and C++ code to the memory-safe language Rust. The benchmark numbers are impressive, with Google claiming that Argon achieves 77.9% on DeepSWE v1.1 and first place on the Vals Index of economically weighted knowledge work.

The model was trained to find, validate and patch vulnerabilities on its own, and will be given to trusted defenders through the Fairwind Program. However, access is restricted for now, with only a small group of vetted cybersecurity defenders able to use it, along with Google's own teams. The company plans to expand access in the future, starting with paid API customers and Google AI Ultra subscribers.

The model has sparked controversy due to its potential capabilities, including being able to find vulnerabilities that earlier models had missed. However, Google has implemented safeguards to prevent misuse, including refusing help with cyber and chemical, biological, radiological or nuclear attacks, and running under a monitor that watches its chain of thought and halts execution when it oversteps the user's intent.