Sora Is Gone: 5 Alternatives That Still Take Your Face (September 2026)

OpenAI wound Sora down in two stages: the app and web experience were discontinued on 26 April 2026, and the API endpoints stop serving on 24 September 2026. If you built anything on Sora 2 or Sora 2 Pro, you have a deadline.

The feature people actually miss is not text-to-video — plenty of models do that. It is putting yourself in the shot. Here is where that goes.

The five that still work

Prices re-read from the live endpoints on 4 September 2026 — all five are unchanged. The 24 September API cutoff is now weeks away, so this is the last comfortable moment to migrate.

HappyHorse 1.1 reference-to-video is the closest thing to the Sora cameo experience: hand it up to nine photos of yourself plus a prompt, get 1080p in a genuine 9:16 or 4:5 frame with native audio and multilingual lip-sync. It is the only one on this list with the full set of vertical aspect ratios out of the box.

Kling 3.0 Motion Control solves a different problem — it copies the movement of a 3–30 second reference clip onto your character. If what you liked about Sora was describing a scene, this is not it. If what you want is to perform someone else's choreography, nothing else on the list competes.

Wan 2.7 edit-video keeps the original video and swaps the performer, so the background, lighting and camera move survive. Cheapest per second of the group and the least talked about.

MiniMax H3, released 31 July 2026, takes up to nine image references plus reference clips for motion and audio, at 2K. It is new enough that there is no meaningful body of user reports yet — treat it as promising rather than proven.

OmniHuman 1.5 is the audio-driven option: a photo plus an audio track produces a performance matched to the audio. Not a scene generator.

What is not a replacement

Seedance 2.0 and Veo 3.1 both refuse real human likenesses, and the filter sits upstream of every reseller — so "use Seedance through a different provider" is not a workaround. If a tool markets itself as a Sora replacement and runs on either of those, it cannot do the cameo use case.

Seedance 2.5, released late August 2026, does not change that. It takes up to 50 references and produces a genuine 30-second single take, but we have not re-tested the real-face filter on it — the filter sits in ByteDance's inference layer, above the model rather than in it, so we treat it as unchanged until the vendor says otherwise. At roughly $0.462 per output second it is also the most expensive per-second rate in the autumn wave. The same applies to the rest of the autumn releases — MiniMax H3 Max, Gemini Omni Flash 1.1, Wan 3.0, LTX 2.5 — which have no published per-output price yet (fal's pricing API returns the platform GPU-fallback rate), so we quote a ceiling rather than a price; the autumn round-up has the full list. None of them displaces the five above for the cameo job.

Consumer face-swap apps in the $10–15/month bracket do a narrower job than they imply: most are face-only swaps at 720p with a 15-second cap, which reads as pasted-on rather than as a performance.

Migration checklist

1. Note the 24 September API cutoff and inventory anything calling Sora endpoints.
2. Decide which of the two jobs you need — new scene from references, or existing clip with a new performer. They are different models.
3. Re-check prices before committing. Every number above is dated, and the endpoints reprice without announcement.
4. Export anything stored in a Sora account. OpenAI has stated the data goes with the shutdown.