Google Launches Gemini 4 Argon — Defenders First

On September 30, Google DeepMind announced Gemini 4 Argon, its first flagship frontier model in more than seven months.

Headline specs: an industry-leading 1-million-token output limit (up from 64K), tuned for long-horizon work — real-world software engineering, enterprise knowledge work (legal, finance), and cybersecurity defense. Introductory pricing is $2 input / $10 output per million tokens.

Google’s claimed results: ties for first on CWE-bench v1 (68%), a new state of the art on DeepSWE v1.1 (77.9%), and Wiz reportedly used it to find a critical flaw in widely-used healthcare software that earlier frontier models missed. Artificial Analysis scores it around 53 on its Intelligence Index — roughly tied with OpenAI’s GPT-6 Astra, just behind Claude’s 5.5 models. (All benchmarks are Google-reported; independent tables weren’t published at launch, so treat the margins cautiously.)

The rollout is the real story: initial access goes only to vetted trusted cyber defenders through the Fairwind Program — including a version “without cyber guardrails” — while Google participates in the US government’s voluntary pre-release access process. Paid API customers and AI Ultra subscribers come next; no general date announced. Thousands of Google employees already use it internally.

Why it matters

A Google flagship always resets the frontier conversation — but the “defenders first” gating is itself the news. Capabilities have crossed a threshold where the company treats cyber-offense potential as a release-management constraint. It sharpens the asymmetry of the AI security landscape: defenders get the tool before the public does.

Good reads

• Introducing Gemini 4 Argon — Google DeepMind blog (Sept 30). The primary source: specs, the Fairwind rationale, pricing, and rollout framing in Google’s own words.
• Gemini 4 Argon: Google’s new model goes to defenders first — The Nano AI (Oct 2). Best analytical take, with independent benchmark context.
• Can Google’s Gemini 4 Argon Catch OpenAI and Anthropic? — CoinCentral (Oct 2). Competitive positioning against Claude 5.5 and GPT-6.1 Sol.
• Gemini 4: Release Date, Argon Launch & Rumors — felloai (Oct 1). Useful roundup including rumored LMArena sightings and pricing tiers.

Elsewhere this week

• Anthropic’s IPO prospectus leaked (Reuters/FT, Sept 29): reported ~$42B 2025 net loss, $518B in planned infra spend, and a possible mid-November IPO targeting up to $2T. Treat as unconfirmed — the figures come from press reports of a confidential document, not a public S-1. (TechXplore summary)
• Anthropic Sonnet 5.5 shipped Sept 28 with frontier-grade cyber safeguards; DeepSeek-V4.1-Flash dropped Oct 4.