Argon targets long-horizon work; a Sume Format run ends at 90 minutes

A model built for long-horizon work does not extend a Sume Format run's clock. Google's Gemini 4 Argon announcement frames the model around long tasks in software, knowledge work and security (read 2026-10-04). Sume's Runs and results page says a run is force-finalized as failed 90 minutes after it was created, or sooner if it is older than 25 minutes and has been silent for 10.

So the ceiling is Sume's, and it holds whichever orchestrator you pick. Plan jobs that fit it.

What are the two clocks?

Both are carried on the status response as expires_at.

How do I split a long job?

Make each run produce one deliverable and chain them in your own code.

• One run per video, not one run per campaign.
• Pass the earlier run's output URLs to the next as input or attachments.
• Use a spend cap on each run, so one failure costs one cap.
• For many similar items, see the bulk docs rather than one giant run.

When Argon opens, will this change?

Only if Sume changes the rule, which the docs would say. A smarter planner might make the same work finish faster, which is the useful result: run the same input on two orchestrators and compare how long each took.