Skip to content
NewA calmer, faster ZeroHour
StoryModel releaseinfo

Google limits Gemini 4 Argon to trusted defenders

3 sources · 3 articlesFirst reported Updated ()2 · 1 read
What's new

No material additions; the prior summary already incorporated all ten reports, including Google's October 2 recap.

Merged summary · grok-4.7 · rewritten as coverage arrivesSummaries by Vani AI

Google is limiting Gemini 4 Argon to Fairwind defenders, citing a 1 million-token output, benchmarks, and a U.S. review.

Google DeepMind announced Gemini 4 Argon, a frontier model for coding, enterprise knowledge work, and defensive cybersecurity, with early access for trusted cyber defenders through the Fairwind Program during a voluntary U.S. government pre-release review. Sources disagree on access: Latent Space also includes government users, while SecurityWeek and Help Net Security say internal teams receive a version without usual cyber guardrails so the model can find, validate, and patch critical flaws, which Help Net Security says it can do without human help. Most outlets say the output limit rises from 64,000 to 1 million tokens—Latent Space credits Long Decode Continuation and calls that an industry first—while Google's October 2 September recap instead describes a 1-million-token context window. Introductory pricing is widely $2 per million input tokens and $10 per million output tokens before doubling to $4 and $20, though Latent Space calls $4/$20 the standard rate with a 50 percent introductory discount and Hacker News adds a 95 percent cached-input discount. Cited scores include 77.9 percent on DeepSWE v1.1, 51.3 percent on Zapier's AutomationBench, 91.7 percent on LVBench, 68.9 percent on the Vals Index, and 68 percent on Collinear AI's CWE-bench v1, where SecurityWeek reports a tie with GPT-6 Astra and Grok 4.7 and CSO Online also includes Claude Opus 5.5; Latent Space says Argon leads 13 of 19 benchmarks against GPT-6 Astra and Claude Opus 5.5 and that Artificial Analysis scores it 53, tying Astra, while CSO Online says Opus 5.5 still leads PostTrainBench at 49.3 percent versus Argon's 45.3 percent. Several outlets say Wiz used Argon to find an unnamed critical personal-data exposure in worldwide hospital software that earlier models missed, DeepMind says internal runs freed over 300 TiB with up to 1 PiB projected including Rust rewrites of re2, libgav1, and more than 800,000 Fuchsia Zircon kernel lines, and a wider release is described as adding cyber and CBRN refusals, prompt-injection defenses, and chain-of-thought monitoring; The Verge says the launch followed OpenAI's Dots agent, GPT-6.1 Sol, and a withheld GPT-6.1 Astra—unlike outlets benchmarking GPT-6 Astra—and Google's recap also lists Fairwind-limited Gemini 3.8 Flash Cyber plus Gemini 3.8 Flash, Live voice models, expressive TTS, Lyria 3.5, and Connected Apps.

Key facts
  • Google DeepMind announced Gemini 4 Argon on 2026-09-30 for coding, enterprise knowledge work, and cyber defense, with early Fairwind Program access for trusted defenders during a voluntary U.S. pre-release review.
  • Most outlets say the output limit rises from 64,000 to 1 million tokens; Latent Space credits Long Decode Continuation, while Google's 2026-10-02 recap instead describes a 1-million-token context window.
  • Introductory price is widely $2 per million input tokens and $10 per million output tokens before $4 and $20; Latent Space calls $4/$20 standard with a 50% intro discount, and Hacker News adds a 95% cached-input discount.
  • Cited scores: 77.9% DeepSWE v1.1, 51.3% Zapier AutomationBench, 91.7% LVBench, 68.9% Vals Index, and 68% on Collinear AI CWE-bench v1.
  • CWE-bench ties differ: SecurityWeek names GPT-6 Astra and Grok 4.7; CSO Online also includes Claude Opus 5.5. Opus 5.5 leads PostTrainBench 49.3% to Argon's 45.3%; Artificial Analysis scores Argon 53, tying Astra.
  • Wiz reportedly used Argon to find an unnamed critical personal-data exposure in worldwide hospital software that earlier frontier models missed.
  • DeepMind says internal runs freed over 300 TiB, with up to 1 PiB projected, including Rust rewrites of re2, libgav1, and more than 800,000 Fuchsia Zircon kernel lines.
  • A wider release is described as adding cyber and CBRN refusals, stronger prompt-injection defenses, and chain-of-thought monitoring that can halt a run; vetted Fairwind access omits usual cyber guardrails.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

Coverage

oldest first · one row per article · UTC
  1. · Sep 30, 2026 Google DeepMindFirst report

    Gemini 4 Argon: our next era of frontier intelligence

    DeepMind unveiled Gemini 4 Argon, a phased frontier model aimed at coding, knowledge work, and cyber defense.

  2. · Sep 30, 2026 Hacker News · AI

    Gemini 4 Argon

    Google launched Gemini 4 Argon for trusted defenders, with 1M-token outputs and defensive cyber capabilities.

  3. · Sep 30, 2026 Ars Technica · AI

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

    Google announced Gemini 4 Argon, a frontier model not yet public, claiming leading coding and cybersecurity scores.