Which model runs a Sume Format? The model field on a run

The optional model field on a Sume Format run selects the LLM that orchestrates the run, and omitting it runs the default, gpt-6-sol. It does not choose the image, video or audio models: the Format's own tools choose those. The receipt echoes the id that actually ran.

The rules are from Create a run, Runs and results and Errors and spend, read 2026-09-29.

What exactly does model select?

The docs define it as an Agents catalog id for the LLM that orchestrates the run. The orchestrator is the LLM that runs the Format's agent turn. The images, video and audio a run generates use models the Format's tools pick, so changing model does not switch your video model.

How do I send it?

Add it to the create body next to the other fields. Most integrations never need it; leave it out unless you have a reason. Send an Idempotency-Key on every create, whether or not you set model.

curl -sS -X POST "https://api.sume.com/v1/formats/acme/promo/runs" \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: promo-model-1" \
  -d '{
    "instruction": "One 9:16 clip",
    "model": "gpt-6-sol"
  }'

When should I set it at all?

Most calls do not need it, since omitting it runs the default. Set model when you want a run pinned to a specific orchestrator, for example to keep a batch on one id while you compare receipts. Because an unknown id is a 400 invalid_request, validate the id against the Agents catalog before you hard-code it in a job.

Note the retired gpt-5.6-sol id: an older script that still sends it keeps working, and the receipt shows it ran on gpt-6-sol.

How do I see which model ran?

Read model on the run receipt from GET /v1/format-runs/{run_id}. The docs describe it as the catalog id the orchestrator ran on, which is the way to confirm what your request resolved to, for example after sending a retired id.

Does the model change what a run costs?

The docs do not tie model to a price. What they state is narrower: usage.billable_amount_usd_micros is generation spend against the cap and excludes the agent's own LLM turn, while usage.debited_usd_micros is what the wallet actually deducted, the turn's own LLM row included. Compare those two on a receipt rather than assuming.

To choose the video model itself, see sume/auto or a pinned video model. For how a Format run's fields differ from a scheduled run's, see the field comparison.