Sora 2's API ends 24 September 2026: where to route
App already gone 26 Apr. API shutdown 24 Sep. Route dialogue to Veo 3.1, references/30s to Seedance 2.5, motion to Kling 3, talking shots to Happy Horse.
Do not start new work on Sora. The app and web experience already died on 26 April 2026. The API dies on 24 September 2026 — thirty days from this post. OpenAI's own discontinuation help is the authority: two dates, no grace generate after the second one, export what you still need.
This is not the eulogy. The history sits in Sora 2 shutdown aftermath and the earlier API sunset migration plan. This page is the 30-day routing table. Sora was a bundle of jobs wearing one name. Split the bundle.
What actually shuts on the 24th
Anything still calling the Sora 2 API — scripts, agents, scheduled jobs, storyboard batches — stops. OpenAI's help page does not describe a fallback response. The route goes dark.
The consumer surface is already gone. If you were clicking around in the app, you already migrated or you already stopped. If you still have a pipeline on the API, you have a calendar problem, not a model-taste problem.
Existing Sora files you care about should be exported now, not "after we pick a replacement." OpenAI points export at sora.chatgpt.com/sunset. Do that in parallel with the routing work. A replacement model will not resurrect a clip you forgot to download.
Four jobs Sora was covering, four models that cover them
Sora-shaped work was never one spec. It was dialogue, a longer take with identity, motion, and a talking shot. Those already live as separate catalog rows. Point each job at the row. Do not hunt for a twin.
| Sora-shaped job | Route to | Why this row |
|---|---|---|
| Spoken line, always-on audio, 4K hero | Veo 3.1 | Audio on/off, 4 / 6 / 8s, 4K listed |
| Identity lock, references, ~30s single pass | Seedance 2.5 | I2V + R2V, durations to 30s |
| Motion value, multi-shot energy | Kling 3 | Cuts and camera, not a presenter |
| Talking shot from a still, multilingual mouth | Happy Horse 1.1 | Native audio and multilingual lip-sync at 1080p |
Read the table left to right, not as a ranking.
Dialogue. If the brief is a line that has to be said, Veo is the router. Native 48 kHz audio, a duration set that looks like a talking shot, 4K when the master needs it. Omni Flash can win an arena week on pretty 8-second audio clips; that is not this row.
References and 30 seconds. If Sora was how you held a character or a product across a longer take, Seedance 2.5 is the replacement that actually publishes that contract: image-to-video from a locked still, reference-to-video with a 50-slot budget you will almost never fill, one pass up to 30 seconds. Flux 3's 20-second native-audio take is a cousin, not the 30-second lock.
Motion. If Sora was how you got camera energy and cuts, Kling 3 is the motion engine. Do not send a talking presenter here and do not send a pack shot you cannot afford to restage.
Talking shots. If the deliverable is a person speaking to camera — especially from a still, especially in more than one language — Happy Horse 1.1 is the specialist. Veo still owns wides, weather, and the cinematic talking hero. Happy Horse owns the talking product explainer. Mixing those two is how you spend Veo credits on a job that never needed 4K weather.
No single row restores Sora's storyboard mode as a product. Multi-scene is an editor problem: generate each beat on the row that owns it, then cut. That is a feature of being off a single-vendor API, not a downgrade.
A 30-day sequence that actually finishes
Treat 24 September as a hard cut, not a feeling.
This week. Freeze new Sora prompts. Export. List every live job that still calls the API and tag it with one row from the table. If a job cannot be tagged, it was several jobs. Split it.
Next two weeks. Rebuild the tagged jobs on the new rows. Do not rebuild them on one "Sora replacement." A talking ad that used Sora because Sora spoke should come up on Veo or Happy Horse. A product take that used Sora because Sora held the bottle should come up on Seedance. A motion bumper should come up on Kling.
Last week of the window. Run the scheduled jobs against the new endpoints until they survive a full cycle. Then delete the Sora route so nothing silent-fails on the 24th.
The point of an aggregator bill is that this sequence is a config change. The point of living in one app is that this sequence is a rebuild. Sora is the case study. It will not be the last one.
FAQ
Can I keep generating on Sora until 24 September?
You can, on the API, until it is discontinued. You should not start new work there. Thirty days is enough to migrate. It is not enough to finish a campaign you begin on a dying endpoint.
What is the closest one-model replacement?
There is not one. Dialogue goes to Veo 3.1, references and 30-second identity to Seedance 2.5, motion to Kling 3, talking shots to Happy Horse 1.1. Hunting for a twin is how people stall until the 24th.
Does ChatGPT still generate Sora clips after the API dies?
OpenAI's help page is about the app already being gone and the API shutting on 24 September. Do not plan production on a residual ChatGPT feature. If it exists, it is not an API, not a batch path, and not a date you control.
I already have the sunset migration plan — is this a duplicate?
No. That post is the history and a wider replacement map. This post is the 30-day routing table for the four jobs Sora was covering. Use both: export and timeline from the plan, job-to-model from here.
The discontinuation notice will not grow a fifth date. Route the jobs this week. Leave Sora off the new-work list.