Which Sume errors mean switch the video model, and which mean wait

Only a few Sume errors justify switching to a different video model. A 404 model_not_found means the id you sent is not in the catalog, so a fallback id is the right move. Most other errors, such as 429 queue_full, 503 provider_capacity_exceeded, or 402 insufficient_credits, say nothing about the model, and switching models would hide a problem that follows you to the next one.

The distinction matters when a vendor removes an API. OpenAI's deprecations page lists the Videos API and the sora-2 family with removal on 2026-09-24 and no replacement named (read 2026-10-08). A fallback chain built after that date should switch on the error that actually means the model is gone, not on every failure.

The decision table

Codes and meanings come from the errors page and the generation admission page. The third column is a recommendation for a fallback chain, not a Sume rule.

Why most errors should not trigger a switch

A queue-shaped error tells you about your workspace, not about the model. queue_full means the workspace has used all of its accepted generation capacity. Moving the same job to another model id would still land in the same queue. Sume's docs describe concurrency as a plan limit, with Free at 1 processing slot and 5 queue slots, which gives 6 accepted jobs.

A 402 is also workspace-wide. The reservation is the estimated cost of the request, so a cheaper model could pass where an expensive one failed. That is a billing decision, though, and your code should ask for it explicitly instead of making it silently.

A small fallback policy

Read retryable and next_action on a failed job before you resubmit. Those fields exist so that the client does not guess.

• Keep an ordered list of model ids in config, not in code.
• Switch only on 404 model_not_found, or on a failed job whose error category you have decided to treat as model-specific.
• Send a new Idempotency-Key when the payload changes, because a reused key with a different body returns 409 idempotency_conflict.
• Log the request id from the error envelope with every switch, so support can follow the path.
• Put sume/auto last only if you accept that Sume does not disclose which family served the request.