Live-commerce run spend cap: $120 typical, $400 default, $500 maximum

For a live-commerce Format run on Sume, set generation_spend_cap_usd to about $120. A Format that never named a cap defaults to $400, the platform maximum is $500, and sending 0 or a number above 500 returns a 400. Sume's calling guide (read 2026-10-06) says production live-commerce integrations run with caps of approximately $120, and that a single-scene retry on the same thread needs a fraction of that.

The cap is a ceiling on one run's generation spend, not a price. The receipt reports the effective cap and the actual spend side by side, so you can tune it from real runs instead of guessing.

The rules in one table

The cap logic is short. If you send nothing, the run inherits the Format's cap. If you send a number up to 500, Sume uses it, even when it is above the Format's own cap. If you send null, you get the platform maximum of $500, which lifts the ceiling without removing it.

Reading the receipt

Every receipt carries usage.generation_spend_cap_usd_micros, the effective cap, and usage.billable_amount_usd_micros, the run's actual spend. After ten runs, compare the two. If the biggest run used a third of the cap, you can tighten it. If a run came close to the cap, raise it for that job instead of lifting every run to $500.

A low cap is protection, not savings: Sume's guide says a run can never spend more than its own effective cap. That is why 0 is refused: a run that cannot spend cannot deliver. Set the cap high enough for the work you want delivered.

Choosing a number

Start at the documented working figure of about $120 for a full host-video run, then size retries separately. A retry of one scene on the same thread needs a fraction of that, so cap retries tighter than first runs. Keep the platform default of $400 for Formats you have not measured, and treat $500 as the break-glass value.

One limit to remember: service-account keys cannot create Format runs at all, and the key needs the formats:write scope. A team workspace's Format needs a key issued in that workspace, so cap questions often turn out to be key questions.