Stripe Issuing for agents vs a Sume spend cap: what limits an agent?
Stripe Issuing caps what an agent can charge to merchants. Sume's required generation_spend_cap_usd caps one run's media spend. They guard different layers.
They limit different things, so a serious agent setup uses both. Stripe's Issuing for agents gives an agent a virtual card with spend limits and merchant controls, which bounds what it can buy from other businesses. Sume's generation_spend_cap_usd bounds how much one Agent Completion can spend on generation inside Sume. One cannot stand in for the other.
Two layers
| Question | Stripe Issuing for agents | Sume Agent Completion |
|---|---|---|
| What is capped | Charges on a virtual card; limits per authorization or over a rolling period, plus allowed merchant categories | Generation spend for one run |
| Where it is set | At card creation, editable later | On every request; no default, a missing value returns 400 invalid_request |
| Live decision | issuing_authorization.request webhook, 2-second default timeout before the card's default rules apply | None per call; the cap is fixed before the run starts |
| Who gets approached | Stripe application or contact required to start | Any key with agent_completions:write |
Why Sume's cap is mandatory
Sume's docs explain that the chat UI protects you with an interactive spend-approval prompt, and a backend caller never sees it. The cap replaces that prompt. You pick a number per run that you accept losing, and the run cannot spend beyond it.
Stripe's page also lists single-use cards, which are scoped to one task and invalidated after use. That is the closest card-side match to a per-run cap, and it suits purchases made through a browser. For purchases the agent never makes on a web page, the per-run cap on the Sume request is the only control that fires before the work begins.
What the cap leaves out
Read the receipt before you treat the cap as a total bill. usage.billable_amount_usd_micros counts generation spend only. The agent's own language-model turn bills to the separate Agent wallet, so a cap of two dollars is not a promise that the whole run costs two dollars. If you need a hard ceiling on everything, set it where each cost lands: the cap on the Sume call, and spend rules on whatever card or wallet pays for the rest.
A sensible stack
Stripe's page notes settlement can take one to two days after the authorization, which is one reason to keep a second, immediate limit in front of the work itself.
- Give each agent its own Stripe virtual card if it buys from outside merchants, limited by category.
- Send
generation_spend_cap_usdon everyPOST /v1/agent/completions, sized to the single task. - Send an
Idempotency-Keyso a retried request returns the original receipt withidempotency_hit: truerather than a second run. - Alert on
outcomeandusagein the run webhook, not on card activity alone.
Sources
More in Pricing
- Sume STT reserves 1 cent by default: send duration_seconds (1-600)
Sume STT lists $0.01 a minute, assumes one minute if duration_seconds is missing, and caps a job at 10 minutes (10 cents). Send the real length.
- Cost of 1,000 AI images on Sume: Soul, Flux 2 Pro, Grok, Nano Banana 2
A thousand single-image jobs cost $10 on Higgsfield Soul, $40 on Flux 2 Pro, $30 on Grok Image and $100 on Nano Banana 2, from Sume's catalog prices.
- Sume Scale $400 includes $800 of usage: fee per usage dollar by plan
Pro pays $1.00 of fee per $1 of included usage, Startup $0.86, Scale $0.50. What that means for a brand that spends the whole allowance each month.
- Sume text to speech price ladder: 500 to 20,000 characters
Sume TTS bills 3 cents for 500 characters, 5 for 1,000, 24 for 5,000 and 95 for 20,000, the per-job cap. Longer scripts split into 20,000-character jobs.
Written by Sume