Muse Video ranks fifth and nobody can use it
Meta's Muse Video sits fifth on Arena's text-to-video board and has never shipped. A four-question filter that strips unusable models out before you plan.
Meta previewed Muse Video on 7 July 2026. It shares Muse Image's pretraining base and adds native audio. Public access was promised "soon to creators and in Meta AI." Six weeks later it still has not been released, and it sits at number five on Arena's text-to-video board with a score of 1457.
That is a model ranked above MiniMax H3, above HappyHorse 1.0, above Sora 2 Pro and above both Veo 3.1 audio entries, which you cannot use, cannot call, cannot sign up for, and cannot put in a plan.
It is not the only one on that board.
Three of the top ten fail an availability check
Arena's text-to-video board as of 14 August 2026 carries 616,845 votes across 45 models. Here is the top ten with an availability column added:
| # | Entry | Score | Can you build on it today? |
|---|---|---|---|
| 1 | gemini-omni-flash | 1512 | Yes |
| 2 | flux-3-video | 1494 | Yes |
| 3 | dreamina-seedance-2.0-720p | 1482 | Yes |
| 4 | dreamina-seedance-2.5-720p | 1477 | Contested — consumer-first launch, endpoint reports conflict |
| 5 | muse-video | 1457 | Never released |
| 6 | minimax-h3 | 1453 | Yes |
| 7 | happyhorse-1.0 | 1428 | Yes |
| 8 | sora-2-pro | 1364 | API shuts 24 Sep 2026 |
| 9–10 | veo-3.1-audio | 1364 / 1363 | Yes |
Three of the top ten fail, and each one fails for a different reason. Seedance 2.5 shipped on 31 July into Jimeng AI and Doubao Pro, which are consumer surfaces rather than APIs, and the reporting on when a developer endpoint followed does not agree with itself — that is a fail-until-you-verify, not a confirmed no. Muse Video has been demoed and nothing more. Sora 2 Pro is being switched off: OpenAI announced the discontinuation on 24 March 2026, the app shut on 26 April, and the API shuts on 24 September 2026, roughly five weeks from now.
If you built a shortlist by reading down that board, you would have a one-in-three chance of picking something you cannot ship on, or cannot confirm you can ship on without checking the endpoint yourself.
Why boards list models you cannot buy
This is not a defect in Arena, and the fix is not to expect the board to change.
A preference board's job is to measure which outputs people prefer. Restricting the roster to commercially available models would make it a worse instrument for that purpose, and it would systematically exclude exactly the frontier releases the board exists to track. Muse Video scoring at 1457 is real information about what Meta's model does. It is simply not information about what you can do this month.
The same logic explains the Sora 2 Pro entry. It earned that score across a long stretch of matchups while it was live. The board records history accurately. Your roadmap does not care about history.
So the responsibility for the availability filter sits with the reader, and it is worth building it as a habit rather than a one-off, because the Elo rating will always be the most visible number on the page and availability will never be a column.
The four-question filter
Run these before a leaderboard entry is allowed into a shortlist. Any "no" removes it.
1. Is there a documented, callable endpoint? Not a waitlist, not a consumer app, not a demo page. A named endpoint with documented parameters, confirmed by you rather than by a secondary source. Seedance 2.5 is the live example of why the second half of that sentence matters: it launched into consumer surfaces on 31 July, and coverage of what happened to developer access afterwards points in more than one direction. Wan 3.0 is the simpler case — public beta from 6 August with a full API described as "soon," which is a documented no.
2. Is it reachable from where you operate? A model live in one region is not live in yours. This one is easy to check and easy to forget, and it is the failure that shows up latest, usually after a pipeline is already built.
3. Is it out of preview? Preview access is real access with undocumented limits. Rate ceilings can change without notice, terms can change, and the whole thing can be withdrawn. Meta demonstrated the withdrawal case directly: it launched a Muse Image feature on Instagram and pulled it on 10 July 2026, three days later, after SAG-AFTRA objected to public accounts being auto-enrolled into a system that let anyone generate images referencing their photos. Meta said it "missed the mark." Muse Image itself stayed live, but the feature people were building around did not survive a working week.
4. Will it still exist in ninety days? Any announced sunset is disqualifying for new work. Sora is the current example and it will not be the last one. If you have anything still calling that API, the migration plan is the urgent item on your list, not model selection.
There is a fifth question that catches a subtler problem: does the entry name match a configuration you would actually use? Board entries carry suffixes for a reason. dreamina-seedance-2.0-720p is a 720p configuration. veo-3.1-audio is an audio variant listed separately from other Veo entries. A score attached to a 720p text-to-video config tells you very little about the same model at 4K with references attached.
What the filtered board looks like
Apply the filter and the top ten becomes seven, in this order: Gemini Omni Flash, FLUX 3 video, Seedance 2.0 720p, MiniMax H3, HappyHorse 1.0, and the two Veo 3.1 audio entries.
That list is shorter, less exciting, and dramatically more useful. Every entry on it is something you can generate with today:
- Gemini Omni Flash video takes a prompt plus optional reference images, source clips, character IDs and audio IDs, at 720p, 1080p or 4K, 4 to 10 seconds.
- FLUX 3 text-to-video generates with native audio directly from a prompt, up to 1080p, 5 to 20 seconds, eight aspect ratios.
- Seedance 2.0 is the callable half of the Seedance pair on that board.
- MiniMax H3 does 2K from text alone, 5 to 15 seconds, six aspect ratios.
- Happy Horse 1.0 text-to-video reaches 1080p at 3 to 15 seconds with native synchronised audio, Foley and multilingual lip-sync.
Notice what the filter did to the ordering question: it removed it. Once you are down to five or six callable models with meaningfully different duration ceilings, resolution ceilings and reference support, the deciding factor is which one fits your brief, not which one is 30 Elo points ahead. That shortlist is what the best text-to-video models is for, and it is the right altitude to make the decision at.
The habit worth keeping
Availability changes faster than quality does. Seedance 2.5's access story is already moving. Muse Video will presumably release. Sora will definitely stop. None of those movements will be reflected in a leaderboard rank, because rank is not what a leaderboard tracks.
So run the four questions monthly, keep the answers next to the scores, and let the filter do the elimination before quality ever enters the conversation. It removes more candidates than any benchmark will, and it removes exactly the ones that would have cost you a rebuild.
FAQ
How can Muse Video be scored if it was never released?
Boards sometimes score models before public availability. The relevant fact for planning is not how the votes were collected but that no public access path exists, which is what makes the score unusable regardless of how it was produced.
Is Muse Video likely to release soon?
Meta said "soon to creators and in Meta AI" on 7 July 2026 without a date, and nothing has shipped in the six weeks since. There is no basis for a timeline beyond that, and the Instagram feature pullback three days after its launch is a reminder that even shipped Meta features are not automatically durable.
What should I do about Sora 2 Pro specifically?
Treat the 24 September 2026 API shutdown as a hard deadline. Anything still pointing at that endpoint needs a replacement selected and tested before then, not after. Duration, native audio and reference handling are the three axes where the substitutes differ most, so test on those rather than on general output quality.
Does this mean leaderboard rank is useless?
No. It is a good filter for narrowing 45 models to a handful. It is a bad basis for choosing between the survivors, and it says nothing at all about whether a survivor is one you can actually call. Use it for elimination, then decide with your own prompts.