Google announces Gemini 4 Argon AI model, but you can't use it yet

Google has spent the summer talking up smaller Gemini Flash models, but now it’s back at the top end of the stack. Gemini 4 Argon is the new headline model, and Google says it leads on coding, knowledge work, and cybersecurity. The catch is simple: outside users can’t try it yet.

Inside Google, though, Argon is already doing real work. The company says its engineers are using it heavily, and that it tapped “fleet-wide telemetry data” to help save 300 TiB of memory across its data centers. That’s not a demo flourish. It sounds like Google wanted a model that could pay for its own keep before it ever showed up in a product.

And the coding claims are more specific still. Google says Argon agents have been helping migrate C/C++ code to Rust across the company, including thousands of lines in the core re2 and libgav1 libraries and more than 800,000 lines in the Fuchsia OS Zircon kernel. That’s the kind of internal use case that says more than a glossy benchmark chart does, because it points to a model being trusted with ugly, real code.

Google does have charts, of course. On the DeepSWE v1.1 software engineering benchmark, Argon scores 77.9 percent, which Google says tops GPT-6 Astra, Fable 5.1, and Opus 5.5. It also claims an industry-leading result on the Vals Index test for economic analysis. Taken together, the pitch is clear: this is Google’s shot at frontier credibility again, just not for public hands yet.