Positioning and access
Gemini 4 Argon is Google DeepMind’s frontier reasoning model for long workflows in software engineering, enterprise knowledge work and cyber defense. Announced September 30, 2026, it reaches approved trusted defenders through the Fairwind Program. A precise broad-release date remains unconfirmed.
Output allowance and integration details
The official announcementspecifies a 1,000,000 output-token limit. This is distinct from input context. Artificial Analysisseparately reports 1M context, text/image/video/speech input and text output. Those are attributed third-party observations; confirm specifications against the documentation for an approved account.
Checked October 4, 2026: the public Gemini API model catalogdoes not yet list an Argon API identifier. No identifier is guessed here. Rate limits, account regions and available endpoint features await the relevant documentation.
Announced pricing
Google announces introductory rates of $2 per million input tokens and $10 for output, changing to $4 and $20 after the promotion. Cached input receives a 95% discount. The promotion end date is unconfirmed. These announced rates do not establish availability to every account.
Benchmark records with their conditions
Google DeepMind reports DeepSWE v1.1 77.9%, Terminal-Bench 4.0 57.4%, AutomationBench 51.3%, LVBench 91.7% and CWE-bench v1 68.0%. These results use different tasks, tools and verification settings; they are not a success-rate forecast for your work.
The evaluation methodologyuses pass@1 and the highest thinking settings unless otherwise specified. Some results are computed by Google and others come from benchmark providers; competitor numbers generally use provider reports. AutomationBench and AutomationBench-AA are separate records.
Further reading and the model family
Read our editorial Gemini 4 Argon: The Upgrade I Want Is Finished Workor explore the Gemini model family. The cover uses Google’s official announcement artwork.