Worst-case spend of a 100-item Sume bulk queue and its spend caps

What is the most a 100-item bulk queue can spend?

There is no queue-level cap. The ceiling is the sum of the cap on every item. A Format's default generation spend cap is $400, and a run can pass generation_spend_cap_usd up to 500. That value is accepted as sent and is not clamped to the Format's own cap. Sending null means $500, not unlimited, and 0 or anything above 500 is a 400.

So the theoretical worst case for 100 items at the default is 100 x $400 = $40,000, and with every item at 500 it is $50,000. Those are ceilings, not predictions, and your wallet balance is the other limit.

How do the numbers line up?

How do I keep a batch inside a budget?

Set generation_spend_cap_usd on every item to what one row should cost, then multiply by the item count for the ceiling you are willing to risk. Per-item values go in each item of items, since an item has the same body as a single run.

Remember what the numbers measure. usage.billable_amount_usd_micros on a receipt is generation spend only and excludes the agent's own LLM turn; the amount actually debited is usage.debited_usd_micros, and usage can be null while a run is in flight. Micros are millionths of a dollar, so 2,500,000 is $2.50.

• Cap each item at a realistic per-row cost, not the default.
• Run a three-row pilot first and read debited_usd_micros from the receipts.
• Keep the wallet balance above the ceiling you computed, or expect admission to refuse new work.
• Cancel remaining children one by one if the first results look wrong; there is no queue-wide cancel.