Gemini 4 Argon is announced but not public: what to run today

You cannot run Gemini 4 Argon in a production agent yet. Google announced it on September 30, 2026, and says it is rolling out first to trusted cyber defenders through its Fairwind Program, with developers, enterprises and consumers to follow, starting with paid API customers and Google AI Ultra subscribers (read 2026-10-04). Until it opens up, a Sume Format run keeps using the orchestrator you already pick, and nothing about your video, image or audio pipeline has to change.

This post separates what Google has said from what Sume documents, so you can decide whether to wait or ship now.

What did Google actually announce?

The announcement frames Argon as a frontier model for long-horizon work: software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense. The page describes a phased rollout, including a U.S. government voluntary pre-release process.

Which model runs a Sume Format run in the meantime?

Sume's Create a run page says the optional model field is an Agents catalog id for the LLM that orchestrates the run. Omit it and the run uses the gpt-6-sol default. An id outside the catalog is a 400 invalid_request. The docs do not list Gemini 4 Argon, so do not assume it is accepted until the catalog says so.

The field selects the orchestrator only. Image, video and audio models are chosen by the Format's tools, so a new orchestrator does not change which video model renders your clip.

How do I get ready for the day Argon opens?

Treat the orchestrator as a setting, not as part of your recipe.

• Keep model out of your Format body and send it per run, so switching is a one-field change.
• Store the model echoed on every receipt next to the run id, so you can tell which orchestrator produced which result.
• Keep a per-run generation_spend_cap_usd; the cap applies whatever model is orchestrating.
• Run the same input on two orchestrators and compare receipts before you move traffic.

Is waiting a mistake?

Probably not. The expensive part of a video run is generation, and that is billed against your cap regardless of which LLM plans it. Moving to a new orchestrator is a test you can run in an afternoon once it is admitted, using the same call you already send to the Format API.