Google opens Gemini 4 Argon and its million-token output to the public

On October 13, 2026, Google updated its announcement: Gemini 4 Argon, unveiled on October 1 as a “frontier model”, is now available outside the small circle it had been limited to. As Next reports, you can now reach the model through the API in AI Studio and Gemini. It is also open to Google AI Ultra subscribers, including in Antigravity, Google’s agent-building platform, and to Pro subscribers in the Gemini app and in AI Mode in Search. It lands right after Meta’s Muse Wonder. LLM releases have become like buses: you wait for ages, then several turn up at once, each one claiming to be the fastest.

Twelve days in the VIP lounge

When it was announced in early October, Argon was being “rolled out to a group of trusted cybersecurity specialists through our Fairwind program”. In practice, only governments, national cybersecurity authorities, critical infrastructure operators and a handful of hand-picked specialists could use it. Gemini 3.8 Flash Cyber had followed the same path before it. Google also said it was taking part in the US government’s voluntary early-access process, as required by the presidential executive order of June 2. On October 1, Google promised to open it to everyone but gave no date. In the end, it took less than two weeks. By Big Tech’s “coming soon” standards, that is practically early.

A million tokens, this time on the way out

The headline feature fits in a single number: the output limit jumps from 64,000 to 1 million tokens, roughly fifteen times more. Million-token context windows are no longer a novelty, since Anthropic’s Fable 5.1 and OpenAI’s GPT-6 Astra have them too. But both rivals cap their output at 128,000 tokens, so Google can claim answers eight times longer than theirs.

According to Google, the model can therefore “deeply analyze and generate hundreds of thousands of tokens in a single step, allowing it to solve complex problems in one shot”. One practical detail remains: someone will have to read those hundreds of thousands of tokens. For whoever has to review it, a pull request written in one go is less a productivity boost than a cancelled weekend.

Home-field advantage

As with every launch, the announcement comes with in-house benchmarks, and Argon usually comes out ahead of GPT-6 Astra, Claude Fable 5.1 and Opus 5.5. Google claims first place on CWE-bench v1 and on DeepSWE v1.1, which measures long, real-world software engineering tasks. On DeepSWE, Argon scores 77.9%, against 74.3% for GPT-6 Astra, 74.2% for Claude Opus 5.5 and 67.4% for Fable 5.1. Google says its model is particularly good “for specialized coding tasks”.

For everyday code, Next advises you to wait and see: Gemini 3.8 did not exactly shine in its test on the Next extension. A benchmark published by the people who trained the model is like the self-assessment at your annual review. It isn’t necessarily wrong, but nobody ever gives themselves “needs improvement”.

Rust: here be dragons

Google is also showing how it uses Argon internally. Its engineers reportedly use it every day, for routine debugging as well as large-scale code migrations and algorithm design. Google’s chosen example is Argon agents migrating code from C/C++ to Rust. The projects range from libraries of a few tens of thousands of lines, such as re2 and libgav1, to the Zircon kernel of Fuchsia OS, which runs to more than 800,000 lines.

This is thin ice. At Microsoft, a stated goal of eliminating all C and C++ by 2030 with the help of AI drew a barrage of criticism. A full U-turn soon followed.

A million output tokens is enough to rewrite a kernel in a single reply. Now Google just needs a reviewer with a million-token attention span, and a bottomless pot of coffee.