Budget a 20-ad batch before you run it: spend caps on Sume

Cap the spend of a 20-ad batch on Sume at three levels: a generation_spend_cap_usd on each Format run, an estimate with dry_run on the MCP paid tools, and a balance check, since a submit fails with 402 insufficient_credits before provider work starts when the balance cannot cover it. None of these needs a price from this post, because Sume's per-job prices are read from the catalog and receipts.

Sources are Sume's Format bulk runs, MCP tools and gates, and Generation admission pages, read 2026-10-02. The 20-ad figure is from Higgsfield's changelog.

Where do I set a cap on a run?

Format run items take their own generation_spend_cap_usd, and the Format API shows it echoed in the run receipt as generation_spend_cap_usd_micros. In a bulk queue each item binds its own cap, so one expensive ad cannot silently use the whole budget.

How do I preview cost from an agent?

Hosted MCP paid tools take dry_run=true for an admission and cost preview that does not submit the job, and an optional max_spend_usd that is enforced only when you pass it. The docs say to prefer generation_admission_preview or dry_run before expensive bursts, and that ordinary single creates do not need it.

What happens to failed jobs?

For image generation the docs say billing is all-or-nothing: a completed generation is billed in full, a failed or cancelled one is not. Check the page for the surface you use, because other surfaces document their own rules.

What is a sensible batch routine?

Preview one item, run one, read its receipt cost, then multiply by 20 and compare with your cap. Queue 20 ad runs overnight with Format bulk runs covers the queue call.

Write the cap in the same place you write the batch size, so the two change together. A batch of 20 with a cap sized for 10 will stop partway, which is the intended behavior, but you want to see it in a test and not on a deadline. After the first full batch, compare the sum of receipt costs with your estimate and adjust the cap by the difference.