Google Unveils Gemini 4 'Argon': Its First New-Generation Flagship Since Gemini 3

What Happened

Google announced Gemini 4 'Argon' late Wednesday (September 30 ET), anchoring its new Gemini 4 generation of flagship models. Reuters, Barron's, and VentureBeat covered it within the hour. The company calls it its 'most performant model yet', benchmarking it directly against OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5. Alphabet stock rose about 1.6% after hours on the news.

Since the Gemini 3 series in November 2025, Google has not shipped a new flagship; the promised Gemini 3.5 Pro missed its June delivery; and in August, DeepMind's leadership changed — Hassabis stepped down as CEO to become chairman, Jeff Dean left to found a company, and Koray Kavukcuoglu took over. Per Google's blog post cited by VentureBeat, Argon first goes to trusted cyber defenders through the Fairwind Program, while Google participates in the U.S. government's voluntary pre-release model access process; broader availability starts with paid API customers and Google AI Ultra subscribers, with no timeline given.

All benchmark figures come from Google's own tests (18 total): Argon leads outright on 12 and ties on 1; GPT-6 Astra leads outright on 3; Claude Opus 5.5 on 2. Results include DeepSWE v1.1 at 77.9% (Astra 74.1%, Opus 5.5 74.2%), Harvey's Legal Agent Benchmark at 19.6% vs 5.4% and 3.8%, and a Gray Swan indirect-prompt-injection attack rate of 0.7% vs Astra's 8.5%. Astra leads FrontierSWE v2 and Terminal-Bench Science 0.1 by 10.5 percentage points, and Opus 5.5 leads Terminal-bench 4.0 by 9 points.

Key Facts

1. The launch: Late 9/30 ET, Google announced Gemini 4 'Argon', the first model of its new Gemini 4 flagship generation (per Reuters, Barron's, VentureBeat).
2. The benchmarks: Across 18 enterprise benchmarks in Google's own tests, Argon leads or ties on 13 (12 outright + 1 tie); Astra leads outright on 3, Opus 5.5 on 2 (per Google materials cited by VentureBeat).
3. The strengths: DeepSWE v1.1 77.9%, AutomationBench 51.3%, Harvey Legal 19.6%, Vals Finance 65.4%, Gray Swan injection attack rate 0.7%; output ceiling raised to 1 million tokens (per Google's disclosure).
4. The gaps: Astra leads FrontierSWE v2 and Terminal-Bench Science 0.1 by 10.5 points; Opus 5.5 leads Terminal-bench 4.0 by 9 points (same source — Google's own tests).
5. The pricing: Introductory $2/$10 per million tokens (cached input at a further 95% discount) — one-fifth of Astra's listed $10/$50, half of Opus 5.5's $4/$20; rises to $4/$20 after the intro period (per VentureBeat).
6. The rollout: First access only for trusted cyber defenders via the Fairwind Program, alongside U.S. government voluntary pre-release review; no public release timeline (per Reuters).

Context

This is the third flagship-class model priced at $2/$10 this week: Anthropic's Claude Sonnet 5.5 on September 28 ($2/$10, 30% faster), OpenAI's GPT-6.1 Sol on September 29 ($2/$10, billed as near-Astra intelligence), and now Argon's intro pricing at the same $2/$10.

Before Argon, Google went a year without a new flagship, a period that included the 3.5 Pro miss and talent departures. Argon's first targets are enterprise workloads — coding, legal, finance, cyber defense — rather than consumer chat.