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

Ars Technica notes Google skipped a long-promised Gemini 3.5 Pro summer and is jumping to Gemini 4 Argon, which it says offers industry-leading performance in coding, knowledge work, and cybersecurity — but not yet for the public. Internal examples include using fleet telemetry to save 300 TiB of memory and agents migrating C/C++ codebases.

Introductory pricing is listed at $2 per million input tokens and $10 per million output, with cached input deeply discounted, and a much higher 1 million token output limit versus 64,000 previously. Fairwind partners such as Wiz are already using Argon; Google says Wiz uncovered a critical hospital-system vulnerability other frontier models missed, without sharing specifics. Chain-of-thought monitoring is part of Google’s misalignment story.

We’re filing this under “capability claims with receipts still sealed.”

The 300 TiB memory save and hospital vuln anecdote are the kind of concrete wins that make Argon worth wanting. They also arrive without the public detail that would let outsiders judge dual-use risk. A million-token output ceiling is a real workflow unlock for long jobs if the model stays on the rails.

I’m watching Google’s chain-of-thought monitoring language carefully. After summer hacking incidents, every lab is promising instrumentation; few are publishing failure rates.

Useful AI should ship with phased access and published prices. We’ll push for the hospital-class findings to be described well enough that the industry can learn — without turning Argon into an unaccountable black box.

Ars Technica, September 2026, Gemini 4 Argon access and pricing report.