guide
The Sora API shuts down on 24 September. Here is how to migrate.
No official replacement, no Sora 3, and a hard cutoff. What to move to, and why list price is the wrong thing to budget on.
OpenAI is switching off the Sora API on 24 September 2026. If you have anything in production that calls it, that is your hard deadline. There is no extension and no grandfathering. The part that catches people out: no official replacement.
Updated 12 September 2026: this guide now includes our own measured benchmark data for Sora 2, Kling 3.0 and Grok, run through the same fixed 40-prompt battery. Where we say a number is ours, it is measured, not copied.
The timeline, and where you are in it
This did not happen suddenly. It has been telegraphed for six months:
| 24 March 2026 | OpenAI announces Sora will be discontinued |
| 26 April 2026 | The Sora app and web experience go dark |
| 24 September 2026 | The API sunsets. Everything stops. |
OpenAI's own announcement post from 24 March.
We’re saying goodbye to Sora. To everyone who created with Sora, shared it, and built community around it: thank you. What you made with Sora mattered, and we know this news is disappointing.
— Sora (@soraofficialapp) March 24, 2026
We’ll share more soon, including timelines for the app and API and details on…
OpenAI's deprecations page lists the Videos API and the Sora 2 model family as deprecated with that shutdown date. It does not name a recommended replacement, and no Sora 3 has been announced.
Why it is being killed
Not because it was bad. Reported figures put Sora's compute burn at roughly $1 million a day against subscription revenue estimated at under $2 million a month, and usage reportedly peaked near one million monthly active users before sliding under 500,000 by the time the shutdown was announced. Video generation is dramatically more expensive to serve than text or images, and the unit economics never closed that gap.
That is worth sitting with, because it is not a Sora-specific problem. Every model on your shortlist is running the same brutal economics. Pick a replacement partly on whether you believe its provider can afford to keep it alive.
What we actually measured
We ran Sora 2, Kling 3.0 and Grok through the exact same fixed battery of 40 prompts (hands, on-screen text, physics, multi-shot continuity, lip-sync) and scored every clip blind. The full breakdown, with the clips themselves, is on the leaderboard. The short version:
| Model | Overall | What stood out |
|---|---|---|
| Sora 2 | 4.78/5 | Best-in-class multi-shot consistency and native audio, right up to the end. |
| Grok | 4.67/5 | Real native audio with tight lip-sync. Also produced our worst single failure: three separate hands manipulating a knife in one clip. |
| Kling 3.0 | 4.45/5 | Generated zero audio across every lip-sync prompt we ran. Not bad sync: no audio track at all, despite third-party reviews describing it as best-in-class for dialogue. |
That audio gap matters more than the score gap if your Sora pipeline depends on generated dialogue or ambient sound. Check it yourself before you commit. Do not take a review's word for a capability when you can watch a test clip instead.
Lip-sync scores for the three models we ran through the battery.
Choosing a replacement
There is no drop-in. What you pick depends on what you were using Sora for:
| If you needed… | Look at | Why |
|---|---|---|
| Closest overall feature parity | Google Veo 3.1 | The safest default. Natively synchronised audio in a single pass, which most competitors still require a separate step for. Not yet in our own benchmark round. See the caveat below. |
| Multi-shot consistency without Sora's price | Kling 3.0 | Matched or beat Sora on every multi-shot and camera-move prompt we ran, at roughly half the credit cost per generation. Confirm you do not need generated audio first. |
| Generated dialogue/audio on a budget | Grok | The only model besides Sora in our round that produced real, correctly-synced audio. |
| Editing-led workflows | Runway Gen-4.5 | The model matters less than the timeline tooling wrapped around it. Not yet in our own benchmark round. |
| Raw output quality | Seedance 2.0 | Tops several independent quality rankings since launch. Not yet in our own benchmark round. |
About the pricing you will find online
Published per-second pricing for these models contradicts itself across sources. In researching this piece we found Kling 3.0 quoted at $0.09–0.14 per second by one comparison site and at $0.029 per second by another, both published in 2026, both stated as fact. Our price radar reads each platform's published prices twice a day instead.
The bigger point stands regardless of whose number you trust: per-second cost is not what a model costs you. What matters is how many generations you burn before you get a clip you would actually ship. Grok put three hands on one pepper and Kling duplicated slices mid-cut, failures that would have needed a reshoot. A model at half the list price that needs three attempts is more expensive than one at full price that lands first time. We track that number, called attempts to usable, for every model we test.
A migration checklist
- Find every call site. Grep for the Videos API endpoint, not just the string "sora". SDK wrappers hide it.
- Write down what you actually use. Max duration, resolution, whether you rely on generated audio, whether you pass reference images. Most migrations fail on a capability nobody documented internally.
- Check audio explicitly, by watching a clip. We found a frontier model marketed for its dialogue quality that generates no audio at all through the platform we tested. Do not learn this after you have shipped.
- Re-test your prompts. Prompts are not portable. A prompt tuned for Sora will not behave the same elsewhere, and the failure is usually silent: plausible output that is subtly wrong.
- Rebuild your cost model on attempts, not list price. See above.
- Do not leave it to September. Late migrations under a hard cutoff go worse than early ones, every time.
What we are still testing
Veo 3.1, Runway Gen-4.5 and Seedance 2.0 are three of the models most commonly recommended as Sora replacements, and none of them have been through our benchmark battery yet. That is next. When they land, the tables above get updated with measured numbers and the caveats come off. Track progress on the leaderboard or read exactly how we test.
Related: OpenAI Sora 2
Frequently asked
When exactly does the Sora API shut down?
24 September 2026. OpenAI's deprecations page lists the Videos API and the entire Sora 2 model family as removed on that date, with no recommended replacement named.
Is there a Sora 3?
Not as of this writing. OpenAI has not announced one. See our separate piece on what is actually confirmed versus rumored for Sora 3, Veo 4 and Kling 3.5.
Does Kling 3.0 really have the best lip-sync, like reviews say?
Not in our own testing. We ran Kling 3.0 through the same 40-prompt battery as Sora 2 and Grok and it generated zero audio across every lip-sync prompt, despite third-party reviews describing it as best-in-class for dialogue.
What should I migrate to if I was using Sora for generated audio?
Grok was the only model besides Sora in our round that produced real, correctly-synced audio. Google Veo 3.1 is also natively audio-capable but has not yet been through our own benchmark.
Cite this
Free to cite and reuse with a link back. Data is updated as new runs and prices come in, so include the date.
SlopTV. (2026). The Sora API shuts down on 24 September. Here is how to migrate.. Retrieved September 12, 2026, from https://sloptv.co/guides/sora-api-shutdown-migration-guide<a href="https://sloptv.co/guides/sora-api-shutdown-migration-guide">The Sora API shuts down on 24 September. Here is how to migrate.</a> (SlopTV)
Daniel Ochoa: Covers model launches, shutdowns and pricing changes as they happen. Reads deprecation notices for a living so you do not have to.