Spend cap for six 10-second 1080p clips: set $25, not the $400 default

A Format run's generation_spend_cap_usd is a ceiling, not an estimate, and the default is much higher than a small job needs. For six 10-second Gemini Omni Flash 1.1 clips at 1080p, the billed price is $1.88 each, or $11.28. If you want room for one retake of every clip, that is 12 generations and $22.56, so a cap of $25 covers it with a little headroom. A Format that never named a cap reports the platform default of $400, and the maximum you can send is $500.

How the cap works

A run can never spend more than its effective cap. You send generation_spend_cap_usd on the create call to set this run's own ceiling; omitting it inherits the Format's cap, null uses $500, and 0 or anything above 500 is rejected with 400. Every receipt reports the cap as usage.generation_spend_cap_usd_micros and the actual spend as usage.billable_amount_usd_micros, so you can check the headroom afterward. In a bulk queue each child carries its own cap, and the queue has no total limit, so the worst case is items times the per-item cap.

Send it on the request

The cap goes in the run body next to the input. The example sets $25; replace the handle, slug and key with your own.

curl -sS -X POST https://api.sume.com/v1/formats/$HANDLE/$SLUG/runs \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: spot-batch-001" \
  -d '{
    "input": {"brief": "six 10s product shots"},
    "generation_spend_cap_usd": 25
  }'

Choosing the number

Price the happy path, add your measured retake rate, then round up to a whole dollar. A tight cap makes a runaway loop cost at most $25, and if a run hits it the run stops short of its goal instead of silently spending more. Check the arithmetic against the live pricing_skus first, because Google, fal and others change list prices often, and the cap is the one number you set yourself.