10 s Seedance 2 run: $3.78 cap plus $0.02 on Haiku or $2.40 on Astra

Short answer

Ten seconds of Seedance 2 at 720p costs 10 x $0.378 = $3.78 at Sume's list price. Set the run's generation_spend_cap_usd a little above that, since the cap governs the media spend. The agent's own model tokens are outside the cap, per the Format runs docs, and land in usage.debited_usd_micros. For six agent turns of 30,000 input and 2,000 output tokens each, that is about $0.024 on Claude Haiku 5.5 and $2.40 on GPT-6 Astra.

So the same clip can end up debiting $3.80 or $6.18 depending on the orchestrator you picked.

The arithmetic

The turn shape is an assumption: 30,000 fresh input tokens and 2,000 output tokens, no caching, six turns. Vendor rates are those read on 2026-10-08.

Choosing the cap

Seedance 2 accepts 4 to 15 seconds, so ten seconds is within range. A cap of $4.00 leaves $0.22 for a retry of a failed second step, but a retry of the whole 10 s clip would be refused at the cap, which is what the cap is for. The Format docs say a cap of 0 is rejected, a value above $500 is rejected, and null means $500.

The agent's turns are not stopped by the cap. If you expect a long conversation on an expensive orchestrator, check the wallet balance separately.

Which orchestrator

For a one-shot render from a fixed instruction, Haiku 5.5 is enough and costs a rounding error. Astra earns its price on tasks where the planning is the hard part, such as a multi-scene script where continuity matters.

Adding a retry and a reference

Real runs rarely render once. If the first take is rejected and you render a second 10 second clip, media spend doubles to $7.56 and the cap must be at least that. A cheaper first pass at 480p costs 10 x $0.17577 = $1.7577, so a draft at 480p and a final at 720p totals $1.7577 + $3.78 = $5.5377. On Astra, the six turns of planning are $2.40 of the total, about 43 percent of that $5.54 media bill; on Haiku it is under half a percent.

Seedance 2 prices come from the lead-verified package table: 480p $0.17577, 720p $0.378 and 1080p $0.8505 per output second, for 4 to 15 seconds.

If you want to budget before you run, work from the cap. Set generation_spend_cap_usd to the media cost you expect plus a margin for one retry, and leave the agent's token cost to the wallet check described in the related post. For a single 10 second 720p clip, a cap of $4 covers the $3.78 render with a small margin; a cap of $8 covers one retry. A cap much higher than that adds risk without adding any value, because the agent could then spend more on media than you intended.

• Draft at 480p plus final at 720p: $5.5377 media.
• Cap for both: at least $5.54.