Every run follows quote → reserve → run → settle to actual. You cannot start billed generation without a hold.
Dual-meter quote
Preflight prices two meters:
- GPU-seconds — Modal time on the selected class (L40S / H100 / CPU).
- Partner calls — Gemini, Kling, Grok, ElevenLabs, Firecrawl, … by unit.
Reserve is the itemized estimate with headroom. Settlement refunds unused credits. A run that would overrun the envelope parks (for example insufficient funds) rather than spending unbounded credits.
Authoring (classify, enhance, surprise, catalog guide) is not billed as a generation hold.
Hold and settle
- Finish under the quote — unused reserve returns.
- Failed user / cancel — settle to what actually executed; the rest returns.
- Infrastructure / orphan failure — zero-charge for that settle (vendor COGS may still be recorded internally).
- Cancel from the run / card Stop control — same settle-to-actual rule.
Insufficient credits park the group (parked_insufficient_funds) instead of a free latch. Resume retries the hold.
Following a run
Progress streams in chat (pipeline / OUTPUTS) and on Nodes. States worth knowing:
- Waiting for a reference — identity plate required; attach photos and Resume.
- QC rewrite — a closed allowlist of widget patches, then re-run the leaf (not a freeform prompt rewrite).
- Parked — funds, policy, or missing refs; the card shows Resume + Cancel.
Successful upstream leaves are kept; a later Resume does not blindly re-mint everything.
Artifacts
Stills, clips, audio, and the assembled video land in Files (session generation folder). They stay in your library subject to storage quota (10 GiB free / 50 GiB paid). Pin or download anything you need to keep offline.
Provenance (workflow version, node, model) rides the receipt so an output can be traced.
Plans
Current plan prices and included credits are on the pricing page. Credits belong to the organization; members draw from the shared balance.