AI research wiki
ai-industry · 2026-10-01 · Tier 3

Gemini 4 Argon: Google's First Frontier Model in Seven Months

Gemini 4 Argon: Google's First Frontier Model in Seven Months

Sources: The Decoder · The Information · @sundarpichai · @GoogleDeepMind · @arena · @demishassabis repost Raw: The Decoder raw · The Information raw and the 2026-10-01 X feed captures

TL;DR

Google launched Gemini 4 Argon on 2026-09-30 (US), its first frontier model in about seven months, pitched at long-horizon workflows, cyber defense and software engineering. It has a 1M-token output limit. Access starts with select testers; API and paid tiers follow. Independent testing (The Decoder) puts it level with GPT-6 Astra but behind Claude Opus 5.5. The per-token price is well below rivals, but Argon burns more than twice as many tokens per task as Astra, so per-task cost is the honest comparison. Arena's Agent Arena lists Argon (High) at #8 with a $0.62 cost per task, on the Pareto frontier.

Google's internal-use claims

  • Argon agents read Google's server data and freed over 300 TiB of data-center memory, with up to 1 PiB expected.
  • Given quantum software to shrink, it found a version needing 40% fewer qubits and operations than the best published human solution.
  • Argon agents are converting legacy C and C++ to Rust, including an 800K-line OS kernel.

These are vendor claims without methodology.

Why it matters here

  • Token-per-task, not price-per-token. The 09-30 DevDay page recorded GPT-6.1 Sol at a fifth of Astra's price. Argon's low sticker price paired with 2x token use is the same lesson from the other side: compare cost per successful task, the metric the llm-routing page adopted on 09-27.
  • A 1M-token output limit is a KV-cache and decode-time commitment, not just a context feature (see kv-cache).
  • Agents optimizing infrastructure (300 TiB reclaimed) is the first frontier-lab claim of agents cutting hardware need directly, relevant to compute-economics.

Links