1. Home
2. News
3. Google

Gemini 4 Argon is out to cyber defenders only, with paid API and Ultra next

Google's own benchmark table has it leading most rows, but it also prints the ones Argon loses.

By Think Facility · October 1, 2026 · Google AI models Save

Google announced Gemini 4 Argon on September 30, 2026, and almost nobody can use it yet. Koray Kavukcuoglu's launch post says it's rolling out to "a set of trusted cyber defenders" through Google's Fairwind Program, and to Google's own teams.

Everyone else waits for a date Google didn't give. When access widens, the post says, it'll start "with paid API customers and Google AI Ultra subscribers". It doesn't mention a free tier or the free Gemini app at all.

Announced

September 30, 2026

Available now

trusted cyber defenders, Fairwind Program

Next in line

paid API customers and Google AI Ultra subscribers

Introductory price

$2 in, $10 out per million tokens

After that

$4 in, $20 out per 1M tokens

Output token limit

1M tokens, up from 64K

What it'll cost when it opens

The API price is $2 per million input tokens and $10 per million output, with cached input at 95% off. That's an introductory rate. A footnote says $4 and $20 apply once it expires, and the post doesn't say when that is.

So its starting price matches GPT-6.1 Sol, which OpenAI released a day earlier (our Sol brief has the details). The bigger change is length. Argon's output limit is 1M tokens, where the last generation stopped at 64K, and Google says a single run can now produce hundreds of thousands of tokens.

There's nothing to call today. The Gemini API's release notes, read on October 1, 2026, don't list Argon or a model ID for it, and its newest entry is a pair of text-to-speech models from September 22. Our Gemini 4 Argon page will change when that does.

Google's own table, losses included

The Gemini page on Google DeepMind's site sets Argon beside GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5, and it leads most rows. On DeepSWE v1.1, a test of long software engineering jobs, it scores 77.9% to Opus 5.5's 74.2%. Harvey's Legal Agent Benchmark is the widest gap: 19.6% for Argon, 3.8% for Opus 5.5.

Google printed the rows it loses, too. Opus 5.5 gets 66.4% on Terminal-bench 4.0 to Argon's 57.4%, and GPT-6 Astra beats it on FrontierSWE v2, 65.5% to 55.0%. On CWE-bench v1, which measures fixing security flaws, Argon and Astra tie at 68.0%. Every number here comes from Google's own page.

Sundar Pichai, Google's CEO, posted the chart himself at 4:02 pm Eastern, a minute before Google's main account announced the model.

Why it's going to defenders first

Google says Argon can find, validate and patch serious software flaws on its own. For trusted defenders and its internal teams, it's releasing a version "without cyber guardrails". Releasing capabilities like that safely needs a phased approach, the post says. Google is taking part in the US government's voluntary process for pre-release model access, and it'll keep tuning guardrails on early testers' feedback before developers, enterprises and consumers get Argon.

Logan Kilpatrick's post gives the same timing in fewer words.

Meanwhile Gemini 3.5 Pro, which Google hoped to ship in June (our Gemini guide has that history), still carries a "3.5 Pro coming soon" banner on DeepMind's Pro page as of October 1, 2026.