Google Announces Gemini 4 Argon AI Model, Access Limited to Testers

What changed

Google has announced Gemini 4 Argon, a frontier AI model still restricted to trusted testers. The company says it is already using Argon for cybersecurity work, large code migrations and data-center optimization, while paid API users and Google AI Ultra subscribers are due to get access before everyone else.

Argon’s headline specifications are substantial:

• A claimed 77.9% score on the DeepSWE v1.1 software-engineering benchmark.
• A 1-million-token output limit, up from 64,000 in previous Gemini models.
• Testing on C/C++ to Rust migrations, including more than 800,000 lines in the Fuchsia Zircon kernel.
• No announced API pricing or general-availability date.

Why it matters

The important shift is not simply that Google has announced another model. Argon is being positioned as a system that can work through unusually large, extended tasks in one pass. A million-token output limit could make it more useful for sprawling code changes, long technical analyses and other work that normally has to be broken into many smaller requests.

Google says Argon has already helped save 300 TiB of memory across its data centers using fleet-wide telemetry data. It also says the model is helping engineers migrate C and C++ code to Rust. Those are company claims, not independently verified results, but they point to the kind of internal workload Google wants Argon associated with: infrastructure-scale work, not just chatbot responses.

Cyberdefense is the model’s first major proving ground. Partners in Google’s Fairwind Program can test Argon, and Wiz says it used the model to find a critical vulnerability in software used at hospitals worldwide. Google says other frontier models missed the flaw, but has not provided specifics. That makes the result intriguing rather than conclusive.

The safety design is equally central. Google says systems monitor Argon’s chain of thought and can stop it when it moves outside permitted bounds. Whether those controls work reliably remains unverified. For now, Argon is less a finished product than a controlled experiment in whether a powerful model can be useful on sensitive systems without becoming another risk inside them.