Wednesday, September 30, 2026
Tech Beat
Sep 30, 2026, 8:11 PMArtificial Intelligence

Google Unveils Gemini 4 Argon With 1 Million Token Limit, but Access Is Restricted

Google unveils Gemini 4 Argon with a 1 million token limit, top coding scores and cyberdefense trials, but no public access, pricing or release date yet.

Listen to this briefingAudio briefing

Summary

Google unveiled Gemini 4 Argon on September 30, 2026, after promising Gemini 3.5 Pro in June and releasing smaller Flash models over the summer. Google claims leadership in coding, knowledge work and cybersecurity. Internally, Argon used fleet wide telemetry to save 300 TiB of data center memory and migrated thousands of C/C++ lines in the core re2 and libgav1 libraries to Rust, alongside more than 800,000 lines in the Fuchsia OS Zircon kernel. It scored 77.9 percent on DeepSWE v1.1, ahead of GPT-6 Astra, Fable 5.1 and Opus 5.5, and Google says it leads the economic analysis Vals Index.

Argon raises Gemini's output limit from 64,000 to 1 million tokens, but API pricing remains unannounced. Google is initially restricting the model to trusted Fairwind Program partners for cyberdefense because models at this scale require phased release. Google says Wiz used Argon to find a critical vulnerability that could expose personal information in hospital systems worldwide after other frontier models missed it, although no specifics were provided. Following high profile summer hacking incidents and concerns about model misalignment, Google says chain of thought monitoring can halt Argon when it moves out of bounds. Enterprise and consumer access is planned without a timetable, beginning after cybersecurity tests with paid API users and Google AI Ultra subscribers.

Positives

  • Gemini 4 Argon scored 77.9 percent on DeepSWE v1.1, exceeding GPT-6 Astra, Fable 5.1 and Opus 5.5.
  • Argon helped Google save 300 TiB of data center memory using fleet wide telemetry.
  • Argon agents migrated more than 800,000 Fuchsia OS Zircon kernel lines to Rust, plus thousands in re2 and libgav1.
  • The 1 million token output limit increases the previous Gemini ceiling of 64,000 tokens more than fifteenfold.
  • Wiz used Argon to uncover a critical vulnerability affecting personal information in hospital systems worldwide.

Risks & concerns

  • Gemini 4 Argon remains unavailable to the public, with no timetable for enterprise or consumer access.
  • Google has not announced API pricing despite confirming that paid API customers will be among the first broader users.
  • Google provided no specifics about the hospital system vulnerability or evidence supporting its claim that competing frontier models missed it.
  • High profile summer hacking incidents have intensified concerns about model misalignment and the safety of frontier systems.
  • Initial access is limited to trusted Fairwind Program testers because Google considers a phased release necessary at Argon's scale.
Primary sourceAI - Ars Technicahttps://arstechnica.com/google/2026/09/google-announces-gemini-4-argon-ai-model-but-you-cant-use-it-yet/
Read full article
Editorial note: Tech Beat summarizes and analyzes third-party reporting. The source link is the authoritative article. This page does not reproduce the full source text.

More From The Wire

Artificial IntelligenceSep 30

Google Unveils Gemini 4 Argon With 1M-Token Output and Frontier Cyber Skills

Artificial IntelligenceSep 30

Meta Muse, OpenAI Dots and Instinct Face Consumer AI’s Harsh Economics

Artificial IntelligenceSep 30

Google AI Search Publisher Payments Draw Backlash Over Tiny Payouts