Google launches Gemini 4 Argon. Cyber defenders get it first, then paying API users

Google unveiled Gemini 4 Argon on 30 September, its new top model for coding, office knowledge work and cyber defence. Most people cannot use it yet. Access starts with security teams in Google’s Fairwind Program, then paid API customers and Google AI Ultra subscribers.

Gemini 4 Argon writes up to a million tokens in one answer

The biggest technical change is output length. Argon can return up to one million tokens in a single response, up from 64,000 in earlier models. Google pitches it at long, multi-step jobs: finding and patching software flaws, legal drafting, financial research and reading charts and long videos.

On Google’s own numbers, Argon scores 77.9% on the DeepSWE coding test and 51.3% on AutomationBench, which checks routine business tasks. VentureBeat counted that it leads or ties on 13 of the 18 benchmarks Google disclosed. OpenAI’s GPT-6 Astra leads on three and Anthropic’s Claude Opus 5.5 on two.

Argon costs $2 per million input tokens at launch

The introductory API price is $2 per million input tokens and $10 per million output tokens. Cached input gets a 95% discount. After the launch period the price doubles to $4 and $20. VentureBeat put the launch price at a fifth of GPT-6 Astra and half of Claude Opus 5.5.

No date yet for Gemini 4 Argon in AI Mode or Workspace

Google said wider release will come “as soon as possible”, without dates for the Gemini app. Its blog post does not mention Search, Workspace or Google Ads. Google also shared the model with the US government before launch under a voluntary process.

New Gemini models have tended to reach Search in steps. Gemini 3.8 Flash, for example, arrived in AI Mode as the third Flash model in six months, and Search Live moved to Gemini 3.8 Live later.

For teams that already pay for model access, the sum is simple. A job that reads 10 million tokens and writes one million costs about $30 at the launch price, against roughly $60 once standard pricing applies.