Gemini 4 Argon announced: restricted defender access before a broader rollout

Google announced Gemini 4 Argon on 30 September 2026, initially for trusted Fairwind Program defenders. Broader paid API and Google AI Ultra access is coming soon.

Its one-million-token allowance is an output limit. Introductory input/output rates are $2/$10 per million tokens, later $4/$20, with a 95% cached-input discount. Check the applicable rate before deployment.

Our assessment: a high output allowance is useful only when the generated work remains coherent and reviewable. For a long report or a substantial code change, check whether later sections agree with earlier decisions. Review references, duplicated work and unresolved assumptions. More space to generate does not tell you how much of the result a team will accept.

Choose a representative task with a clear finish condition before comparing models. Keep the source material and instructions identical, then assess the completed result against your current workflow. Record the time spent reviewing and repairing it. If an ambitious response takes longer to check than the existing approach, that cost belongs in the comparison.

Keep the announced model on an evaluation shortlist until your organisation can use the intended route. Confirm the native identifier, input limits, supported tools and data handling terms from the documentation attached to that route. A model name in a news announcement is not enough information to configure a production client. Avoid filling missing specifications with assumptions from earlier Gemini models.

When evaluating cost, include the complete job: input, generated output, repeated attempts and human review. Check how cached input applies to your request pattern rather than assuming every prompt qualifies. For a service that produces lengthy responses, set a sensible output budget and a clear stopping condition. These are our deployment recommendations; this article reports the announcement without claiming an AZ Labs integration or independent performance test.