$8 cap vs $6 wallet: which gate stops a 12-shot mixed project first
A 60-second project costs $8.22 on list prices. Four wallet and spend-cap pairs show whether a 402 or a failed run stops it, and how far each gets.

A spend cap and a wallet balance are two different gates, and the lower one stops the project. The wallet gate answers with 402 insufficient_credits and nothing runs for that request. The spend cap is a ceiling on one run: when a run tries to go past it, the run ends as failed. Direct API calls have only the wallet gate; a cap exists on Format runs and agent completions.
The project here is 60 seconds: 3 stills at 2K, a 900-character voice-over, a music bed, twelve 5-second wan-3.0 shots at 720p and a one-minute render. Estimated generation total: $8.21775 ($8.22), from the Sume catalog rates (read 2026-10-08).
What each gate does
The Formats error reference separates the two. The wallet is checked at create, and a failure is a 402 with next_action: add_funds; nothing ran. The cap is checked during the run; the run fails and its usage receipt shows how near to the cap the spend got. Spend counted against the cap is reserved plus captured generation rows, and it excludes the agent's own LLM turn.
Money you already spent stays spent. The docs say you pay for generation that finished before a cancel or a failure.
Four pairs, same project
Assume the steps run in this order: stills, voice, music, shots 1 to 12, render. The table walks that order and stops at the first gate that would be crossed. The order is an assumption; an agent may order steps differently.
| Wallet | Run cap | Outcome | Steps finished | Spent before stop |
|---|---|---|---|---|
| $50.00 | $8.00 | cap stop at shot 12 | 14 of 16 | $7.4928 |
| $6.00 | $20.00 | 402 at shot 9 | 11 of 16 | $5.6178 |
| $50.00 | $8.15 | cap stop at render | 15 of 16 | $8.1178 |
| $50.00 | $10.00 | completes | 16 of 16 | $8.2178 |
Reading the result
Row 1: the $8 cap stops the run at the twelfth shot, even with $50 available; eleven shots, 55 seconds, are paid for and the render never runs. Row 2: a $6 wallet with a generous cap fails with a 402 at shot 9 and the cap never matters. Row 3 is the trap: a cap only $0.03 above the twelve shots leaves no room for the $0.1000 render. Row 4 has room for everything.
Size the cap to the whole estimate, render included, plus the retries you accept. Size the wallet to at least the cap, because a cap larger than the balance just hands control back to the 402.
Which one to raise
When a project stops, read the error before you spend more. A 402 means raise the balance. A failed run means compare usage.billable_amount_usd_micros with usage.generation_spend_cap_usd_micros before you raise the brief or the cap, because the run may simply have been priced above its ceiling.
Both fixes cost money, so check what you are buying. If eleven shots are already paid, a new run with a higher cap should continue from those files, not recreate them. A continuation in Formats is a new run with its own receipt and its own cap, so size it for the missing shots and the render only: 1 shot at $0.625 plus $0.10 is $0.725.
Sources
Related posts
More in Developers
- 1,200 SKUs as 12 bulk queues of 100: keys, concurrency, replays
Format bulk runs take 1 to 100 items, so 1,200 SKUs are 12 queues. Use 12 distinct Idempotency-Keys: a replayed key returns 202 with the old queue.
- 200 video slots in a 180 s Short: 0.9 s each on Sume Timeline
Timeline 1.0 takes up to 200 video slots and 1,800 s of audio. In a 180 s Short, 200 slots average 0.9 s, the render bills $0.30, and fades cap at 0.45 s.
- 250 prompts at 12s 1080p after Sora: which Sume models reach it
Four of the six candidate models make a 12-second 1080p clip on Sume, from $420.00 to $2,042.50 for 250 prompts. minimax-h3 and Omni Flash cannot.
- A 28-minute talk: three detach ranges, three STT requests, $0.31
Audio detach outputs up to 900 s but Sume STT takes up to 600 s, so a 1,680-second video needs three ranges: 3 x $0.01 detach + 28 x $0.01 STT = $0.31.
Written by Sume