Dry run a 30-second AI video before paying: Sume MCP generate_video
Preview a 30-second clip's cost with dry_run and max_spend_usd on Sume's hosted MCP before you submit. Which models reach 30 s and what the gates do.
To check what a 30-second clip will cost before you pay, call generate_video on Sume's hosted MCP with dry_run=true. It returns an admission and cost preview and does not submit the job. Add max_spend_usd to the real call so the submit fails if the price is above your ceiling.
Thirty seconds is the length that now matters, and long clips are the ones where a wrong model choice costs the most. These steps follow the MCP tools and gates and video generation docs, read 2026-10-02.
Which models can I preview at 30 seconds?
Only models whose catalog entry lists 30 in supported_durations. Per the docs these are seedance-2.5 (4 to 30 s), wan-3.0 (2 to 30 s), higgsfield-genjutsu (4 to 30 s, source video required) and h3-max-recast (5 to 30 s, source video required). Every other model is 15 s or less, and sume/auto makes 3 to 10 s clips.
What does each gate do?
The gates are separate and all optional except the key.
| Gate | Required | What it does |
|---|---|---|
idempotency_key | Yes, on paid tools | Dedup and transport key, not human approval |
dry_run=true | No | Admission and cost preview, job not submitted |
max_spend_usd | No | Enforced only when you provide it |
generation_admission_preview | No | Separate preview tool, useful before a burst |
How do I phrase the agent instruction?
Tell the agent in one sentence: call tools_schema for generate_video, run it with dry_run=true for seedance-2.5 at 30 seconds, report the preview, and only submit after I say go with max_spend_usd set to that number. The final submit needs a fresh idempotency_key; reuse it only to retry the same request.
A dry run is not a guarantee of output quality. Make a short clip first if the prompt is untested.
What if the preview says no?
A preview can refuse for wallet or admission reasons. Read the error, top up or pick a shorter duration, and run the preview again. Do not loop on a refused submit; the key makes retries safe, not useful.
Sources
Related posts
More in Developers
- Dub a 20-minute video: audio detach 900 s cap, STT 600 s, TTS 1,200 s
A 20-minute video needs chunking before a dub: audio detach outputs at most 900 s, STT reservation tops out at 600 s, and TTS fails past 1,200 s.
- ElevenLabs Music force_instrumental and seed vs Sume prompt-only
ElevenLabs Compose has force_instrumental and seed fields. Sume's Music Router has neither: you write no vocals into the prompt and check the audio.
- ElevenLabs Music output_format list vs the file Sume returns
ElevenLabs Compose lists MP3, PCM, mu-law, A-law and Opus formats. Sume's music request has no format field; get wav through Timeline audio.
- ElevenLabs Music v2.5 became the default: which model made a track?
ElevenLabs made Music v2.5 the default on 2026-09-11. When a default changes, old and new tracks differ. How to record the engine per track, on Sume too.
Written by Sume