Does the Sume spend cap cover the agent's own model tokens?

The answer

The cap does not include the orchestrating model's own tokens. The Format runs docs say usage.billable_amount_usd_micros, which the cap is enforced against, "does not include the LLM turn of the agent, thus it is not the total cost of the run." The real cost is usage.debited_usd_micros, which contains the LLM row.

So a run with a $5 cap can debit slightly more than $5 once the model's tokens are added. The size of that extra depends heavily on which orchestrator ran.

How big is the extra

Take one assumed turn shape: 30,000 fresh input tokens and 2,000 output tokens, no cache. This is an illustration, not a measurement. The rates are the ones in Sume's rate cards, which match the vendor pages read on 2026-10-08.

What to set

Set generation_spend_cap_usd for the media you expect to buy, then leave room in the wallet for the debit that sits outside it. The Format docs give $500 as the platform maximum for a single run. Production live-commerce integrations run around $120, per the same page.

For an Astra run, the wallet also needs headroom at the start: the repo reserves a ceiling on the turn before it runs and releases the unused part. The Astra reserve post covers it.

Where Agent Completions differ

The Agent Completions receipt documents usage.generation_spend_cap_usd_micros in its example. It does not show the debit fields in the sample, so poll the receipt and read usage on a completed run rather than assuming the Format shape.

Practical sizing rules

Treat the cap and the wallet as two different checks. The cap answers how much media this run may buy. The wallet answers whether Sume can start the turn and pay the model. For Haiku 5.5, the model line is a rounding error next to a clip. For Astra it is not: ten turns of the assumed shape cost $4.00, which is more than a whole 10 second Seedance 2 clip at 480p ($0.17577 per second x 10 = $1.7577).

Two habits help. After each run, store usage.debited_usd_micros next to usage.billable_amount_usd_micros; the difference between the two is the part the cap did not see. And when you compare models, compare the debited figure, not the capped one, or the cheaper-looking model may simply be the one that did less thinking.

In short, the cap bounds media, the wallet bounds tokens, and the receipt shows both, so read both fields on every run you intend to repeat.

Re-check the numbers against the rate cards when you budget, because they are a snapshot from 2026-10-08 and can change.

• usage.cap.counted_usd_micros equals the billable amount the cap is enforced against.
• held_usd_micros and refunded_usd_micros are holds, not spend.