benchr
Meta previewed Muse Video on July 7 alongside Muse Image. Native audio is the detail worth watching. “Preview” is the detail that keeps it out of a production comparison.
Put it on a watchlist, not a roadmap
An early preview shows where Meta is heading. It does not promise availability, rate limits, content-policy behavior, commercial terms, or a durable integration surface. Those gaps are too large for a production dependency.
Native audio could remove a workflow step
Generating picture and sound together may be simpler than stitching silent video to a separate track. Meta has not yet published duration, resolution, editing controls, licensing, quotas, or developer access. Keep the claim limited to native audio support.
There is nothing honest to rank yet
No public price table, model ID, billing unit, or benchmark method means no defensible cost, latency, or quality rank. For now, the useful record is one confirmed feature and a list of contract fields Meta still needs to publish.
| Field | Verified record |
|---|---|
| Status | Early preview |
| Capability | Native audio support |
| Public API model ID | Not published in checked announcement |
| Price / limits / benchmarks | Not published in checked announcement |
Native audio needs its own failure checklist
Audio generated with the picture can remove a separate production step, but it also couples two kinds of error. A useful preview review should inspect speech timing, ambient continuity, sudden cuts, off-screen sound, and whether a revised visual scene preserves the intended audio. Meta itself identifies audio-video synchronization and fast motion as current gaps, so those cases belong at the front of the test set.
| Scene | Observe | Why it matters |
|---|---|---|
| Speaking subject | Lip timing, identity, interruptions | Exposes synchronization errors |
| Fast movement | Temporal coherence and sound continuity | Targets a limitation Meta names |
| Ambient sequence | Background audio across cuts | Tests whether native audio survives editing |
| Revision request | What changes and what stays fixed | Measures controllability, not spectacle |
Preview evidence is useful only when preserved
Record the exact prompt, access surface, date, output length, and every edit request. Keep the original files rather than a social-media capture. That record lets the team distinguish a model change from a different prompt or product setting when the preview evolves.
Four gates before a production roadmap
Wait for public availability, a named developer interface, measurable output constraints, and commercial terms. After those gates open, run a rights review and a workload-specific test before estimating cost. A preview can justify discovery work; it cannot justify a delivery date.
Build the comparison around a finished sequence
A single impressive clip hides the hard part of video production: continuity across shots. Write a three-shot sequence with a stable subject, a change of camera position, and a final action. Test whether appearance, motion, lighting, and ambient audio survive the sequence. Keep each generation and failed revision; selecting only the best clip creates a portfolio, not an evaluation.
Score visual continuity and audio continuity separately. Then add a combined review for whether the sound belongs to the visible event at the correct moment. If the preview does not expose deterministic controls, seeds, or a formal editing surface, mark repeatability as unknown rather than penalizing or rewarding a feature that has not been documented.
For a real buying decision, add turnaround time, usable-output rate, review labor, and post-production work after public access exists. The cheapest generation is not the cheapest finished scene if editors repeatedly rebuild timing or sound. None of those operational numbers can be inferred from Meta's preview announcement, so this page leaves them as a future test plan instead of filling a chart with estimates.