The defense-first frontier model, priced like the middle tier. Gemini 4 Argon (September 30) is Google’s new frontier: real-world software engineering, legal and finance knowledge work, and a defensive-cybersecurity lead - rolled out first to vetted cyber defenders through the Fairwind Program while Google iterates guardrails under the US government’s voluntary pre-release process. Broad developer access follows “as soon as possible”; nobody outside Fairwind can run it today.
The pricing datapoint. Introductory $2 per million input tokens and $10 output, cached input at 95 percent off - the same slot OpenAI’s GPT-6.1 Sol opened the day before. Two flagship-class labs pricing their best defensive models at the cheap tier in two days is the month’s thesis datapoint: the frontier competes on price now, and gates its autonomy instead (OpenAI shelved the GPT-6.1 Astra upgrade the same week; Google gates Argon’s access rather than shipping it wide).
Google’s fielded receipts, not benchmarks: Argon agents applied memory optimizations across Google’s own data centers - 300+ TiB freed at rollout - and beat a published quantum-algorithm baseline by 40 percent in minutes. These are internal claims, no external replication, no published eval tables at announcement.
Hosted only. No weights, no local run.