September 30, 2026 — Google DeepMind just unveiled its next flagship model.
What happened
Google DeepMind announced Gemini 4 Argon on September 30, calling it the first model of the Gemini 4 generation — a new frontier release aimed at long-horizon software engineering, enterprise knowledge work like legal and financial research, and cybersecurity defense. The announcement, published on The Keyword by DeepMind SVP and chief AI architect Koray Kavukcuoglu, frames Argon as a new generation rather than an incremental update.
The headline spec: a million tokens of output
Argon’s biggest technical jump is generation length. It can produce up to 1 million tokens in a single response, up from 64,000 on earlier Gemini models. For developers, that means large refactors, long reports, and deep multi-step reasoning chains can run in one continuous trajectory instead of being split across turns. Google says the model set a new state of the art at 77.9% on DeepSWE v1.1, a real-world software engineering benchmark, and scored 68% on CWE-bench v1 for autonomously finding, validating, and fixing software vulnerabilities. Google is already using Argon internally, including on quantum computing algorithm work and large-scale codebase migrations.
A phased rollout — defenders first
You can’t use Argon yet. Google is rolling it out in stages: first to trusted cyber defenders through its Fairwind Program, a controlled-access program for vetted defenders including governments and critical-infrastructure operators, with rules restricting use to defensive and research work and barring sharing or reselling access. Broader access will start with paid API customers and Google AI Ultra subscribers, with no public timetable announced. Google is also participating in the U.S. government’s voluntary process for pre-release model access, and says it will keep hardening safeguards — against misuse, prompt injection, and misalignment — with early testers before widening availability. Alphabet shares rose in extended trading on the news.
Introductory pricing is already public: $2 per million input tokens and $10 per million output tokens, with cached input tokens at a 95% discount. After the introductory period, pricing moves to $4 per million input and $20 per million output.
Why it matters
Google is pairing big capability claims with a deliberately slow rollout — the opposite of the “ship it and see” era. If the million-token output window holds up in real use, it changes what’s practical for agentic coding workflows: entire refactors as one task instead of dozens of babysat prompts. The defenders-first approach is also a tell — this model class is powerful enough that Google wants it finding vulnerabilities for the good guys before it can find them for anyone else.
Sources: Investor’s Business Daily, MarkTechPost, Unite.ai, The Verge