
Sora API: Pricing, Specs, and the Sept 24 Shutdown
The Sora API shut down on September 24, 2026; the app closed in April. Sora 2 specs, Arena rank and live replacements: SeeDance 2.5 and Gemini Omni Flash 1.2.
The OpenAI Sora API stopped working on September 24, 2026. OpenAI's own video guide now opens with a banner saying the Sora 2 models and the Videos API are shut down, with no one-to-one replacement, and the Sora app closed five months earlier, on April 26. A Sora 2 API call today returns nothing usable. What is left to decide is where a Sora pipeline moves.
TL;DR:
sora-2andsora-2-prowere retired together with OpenAI's Videos API on September 24, 2026, after a March 24 notice, and OpenAI names no successor. There is no free or paid route to Sora 2 from OpenAI now. On Arena's text-to-video board (dated November 7, 2026) Sora 2 Pro had already slipped to #13 with 1367 points. For one long take with sound, we would move to SeeDance 2.5, which runs up to 30 seconds per request against Sora's 20. For Sora's extend-and-edit workflow, Gemini Omni Flash 1.2 has both endpoints and leads that board at 1540. Both are live on UnificAlly.
Sora API shutdown: what OpenAI's pages say now
Three OpenAI pages agree, and all three were re-read on November 16, 2026:
| Item | Status | Source |
|---|---|---|
| Sora app and sora.com | Discontinued April 26, 2026 | OpenAI help center |
Videos API (/v1/videos) | Shut down September 24, 2026 | OpenAI deprecations |
sora-2, sora-2-pro | Shut down September 24, 2026 | same table, under "Past deprecations" |
Snapshots sora-2-2025-10-06, sora-2-2025-12-08, sora-2-pro-2025-10-06 | Shut down September 24, 2026 | same table |
| Listed replacement | None ("No one-to-one replacement") | same table |
| Notice date | March 24, 2026 | same table |
In OpenAI's lifecycle terms, "shutdown" means the standard API refuses the model outright, so code still pointing at sora-2 fails instead of falling back to anything. The video generation guide stays online for historical reference only.
Two of the five pages Google ranked highest for "sora api" when we checked still read as if the endpoint works: a forum thread telling people they can generate until September 24, and a third-party page offering "Sora API access". Neither has caught up with the shutdown.
Sora 2 API pricing, and what a 20-second video costs on a live model
Sora API pricing was per second of video, with one rate for sora-2 at 720p and higher rates for sora-2-pro that rose with each resolution step (720p, 1792x1024, 1080p). The Batch API ran the same jobs at a discount. None of that can be bought today, so a search for "sora 2 price" now has one practical answer: what the same video costs on a model you can still call.
Take the longest single Sora generation, 20 seconds at 720p, and rebuild it two ways on UnificAlly:
| Route | How the 20 seconds are made | Cost on UnificAlly |
|---|---|---|
| SeeDance 2.5 | One request, duration: 20, audio on by default | $4.06 |
| Gemini Omni Flash 1.2 | A 15-second generation, then a 5-second extend from its task_id | $0.642 + $0.214 |
| SeeDance 2.5, 480p draft first | Same 20 seconds at 480p to check the idea, then the 720p run | $1.74 for the draft |
SeeDance 2.5 is the simpler bill for long takes because it does in one request what Sora needed an extension for: its duration goes to 30 seconds. Gemini Omni Flash 1.2 prices every whole second from 3 to 15 at each resolution, and its extend endpoint bills only the seconds you add. A task that fails is not billed on either route.
Is Sora 2 free, and do you still need a Sora invite code?
Is Sora 2 free? Not anymore, and not at any price. At launch on September 30, 2025, the app gave free generations with limits and ChatGPT Pro subscribers got Sora 2 Pro on sora.com. That ended on April 26, 2026, and the paid API followed in September.
A Sora invite code is worthless now. The app launched invite-only in the US and Canada, and codes briefly traded hands, but with the app closed there is nothing left to join. Any site still selling a code is selling access to a shut service.
If you bought ChatGPT/Sora credits and never spent them, OpenAI's help center says they can still be used for Codex. Refund questions go through OpenAI's ChatGPT subscription refund process.
Sora 2 Pro on Arena: twelve models now rank above it
Sora 2 Pro was a strong video model with synchronized sound, and Arena still ranks it on old votes. The text-to-video board dated November 7, 2026 (891,840 votes, 50 models) puts it here:
| Model | Arena rank | Score | Callable on UnificAlly |
|---|---|---|---|
| Gemini Omni 1.2 Flash | #1 | 1540 (preliminary) | Yes |
| Gemini Omni 1.1 Flash | #2 | 1513 | Yes |
| Gemini Omni Flash (first release) | #3 | 1509 | Yes |
| FLUX 3 Video | #4 | 1491 (preliminary) | Yes |
| Grok Imagine Video 1.5 | #5 | 1488 (preliminary) | Yes |
| SeeDance 2.0 (720p) | #6 | 1483 | Yes |
| SeeDance 2.5 (720p) | #7 | 1479 | Yes |
| Wan 3.0 | #8 | 1476 | Yes |
| Sora 2 Pro | #13 | 1367 | No, shut down |
| Veo 3.1 (Audio, 1080p) | #15 | 1363 | Yes |
| Sora 2 | #19 | 1342 | No, shut down |
On the November 7 Arena text-to-video board, Sora 2 Pro scores 1367, 116 points under SeeDance 2.0 and 173 under Gemini Omni 1.2 Flash. In August it sat at #8; the board moved past it while it was still callable. The only live model close to it is Veo 3.1, whose four audio variants score 1359 to 1363, inside Sora 2 Pro's margin. The 1.2 Flash score rests on 1,714 votes, so expect it to shift a little; the gap to Sora is far larger than its ±16 margin.
Sora 2 alternatives mapped parameter by parameter
The Sora API had a specific shape: POST /v1/videos, sora-2 for drafts, sora-2-pro for 1080p, 16- or 20-second generations, an extensions endpoint, an edits endpoint and reusable characters. Here is where each of those pieces lives on a model you can call today, taken from the UnificAlly docs:
| Sora 2 API feature | Sora limit | Live replacement | Parameter and limit |
|---|---|---|---|
| Single generation length | 16 or 20 s | SeeDance 2.5 | duration 4 to 30 s |
| Gemini Omni Flash 1.2 | duration 3 to 15 s, any whole second | ||
| Extensions | +20 s per call, 6 calls, 120 s total | Gemini Omni Flash 1.2 Video Extend | task_id + duration 3 to 10 s per call, source made at 360p or 720p |
| Targeted edits | /v1/videos/edits | Gemini Omni Flash 1.2 Video Edit | edits up to 15 s of an existing video |
| First-frame image | input_reference, human faces rejected | Omni 1.2 / Veo 3.1 | start_image_url, plus end_image_url for a last frame |
| Reusable characters | 2 per video, non-human only | Gemini Omni Flash 1.2 | reference_characters, up to 3, each with 1 to 10 images |
| Sound in the same file | always on | SeeDance 2.0 and 2.5 | generate_audio, default true |
| Kling 3.0 | native_audio, default false, billed at the audio rate | ||
| 1080p | sora-2-pro only | Omni 1.2 | resolution 1080p or 4k, upscaled inside the same task |
| SeeDance 2.0 Pro | 720p, 1080p, 4k | ||
| Multi-scene in one request | not offered | Kling 3.0 | multi_shots, 2 to 6 shots, each with its own prompt and duration |
Two differences catch people moving over. Sora blocked human likeness in character uploads by default; the Omni 1.2 docs list no such block, and their own sample character is a person named Mia. And Sora's extensions refused characters entirely, while the Omni 1.2 extend endpoint takes up to 2 character items and 3 reference images for the new part.
Sora's content rules went further than any of these parameters: only material suitable for viewers under 18, no copyrighted characters or music, and no real people, public figures included. Each live model applies its own policy, so re-test prompts that sat near those lines.
Where the other live models fit:
- SeeDance 2.0: 4 to 15 seconds at up to 4K on Pro, up to 9 reference images, 3 reference videos and 3 audio files per request. #6 on Arena at 1483, a score built on 72,137 votes, so it is unlikely to move much.
- Kling 3.0: 3 to 15 seconds, 720p in
std, 1080p inpro, plus a 4K mode, from $0.0756 per second. The only one here that cuts between scenes inside one request. - Veo 3.1: 4, 6 or 8 seconds, 720p native with 1080p and 4K by in-task upscale, from $0.075 per video. Closest to Sora's look on Arena, but the shortest maximum length of the group.
- Grok Imagine: xAI's family; Grok Imagine Video 1.5 sits #5 on Arena on a preliminary score.
Moving a Sora 2 API request to Gemini Omni Flash 1.2
The old call, which no longer runs, sent form fields to OpenAI:
# Sora 2 (shut down September 24, 2026)
curl -X POST "https://api.openai.com/v1/videos" \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-F model="sora-2" \
-F prompt="Wide shot of a red fox crossing a snowy field at dawn" \
-F size="1280x720" \
-F seconds="8"
The same request on Gemini Omni Flash 1.2 is JSON to POST /v1/tasks. size becomes resolution plus aspect_ratio, and seconds becomes an integer duration:
curl -X POST https://api.unifically.com/v1/tasks \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "google/gemini-omni-flash-1.2-video",
"input": {
"prompt": "Wide shot of a red fox crossing a snowy field at dawn",
"duration": 15,
"resolution": "720p",
"aspect_ratio": "16:9"
}
}'
To continue it the way Sora's extensions did, pass the finished task's ID to the extend model with the seconds to add:
curl -X POST https://api.unifically.com/v1/tasks \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "google/gemini-omni-flash-1.2-video-extend",
"input": {
"task_id": "YOUR_COMPLETED_TASK_ID",
"prompt": "The fox stops at the treeline and looks back toward the camera",
"duration": 5
}
}'
Sora handed back a video_id and a download link that expired after one hour. Here the task ID is what you poll (GET /v1/tasks/<task_id>), and the extend call only accepts a source generated at 360p or 720p, so keep any video you plan to extend at 720p or below.
Does Sora put a watermark on videos?
Yes, at least on app output. OpenAI's Sora 2 launch safety notes said every output carried a visible watermark, the moving cloud logo seen across social feeds, and that all Sora videos carried C2PA provenance metadata, backed by OpenAI's internal reverse search for video and audio. OpenAI's help center describes the same approach as Content Credentials plus invisible watermarks. The Videos API docs never described a visible mark on API files.
That question now matters only for videos already made. We don't point to watermark removers: stripping the Sora watermark and provenance data so a generated video passes as real is what C2PA exists to stop, and most platforms require AI disclosure either way.
Sora FAQ
When does the Sora API shut down?
It already did. OpenAI shut down the Sora API, meaning the Videos API and every sora-2 and sora-2-pro alias and snapshot, on September 24, 2026, six months after its March 24 notice. The Sora app and sora.com closed earlier, on April 26, 2026.
Can I still export my old Sora videos?
OpenAI's help center still lists the export: open sora.chatgpt.com/sunset, click Export, and wait for the email. It also says Sora data is permanently deleted once any final export window passes, and that it will email users before such a window starts, so export now rather than later.
How much does a 10-second Sora 2 video cost?
Nothing you can pay today, because OpenAI no longer sells Sora 2 generations. The same 10 seconds at 720p costs $2.03 on SeeDance 2.5 or $0.428 on Gemini Omni Flash 1.2, billed only when the video completes.
Will there be a Sora 3?
OpenAI has not announced one. Its deprecations table lists "No one-to-one replacement" for the Videos API and the Sora 2 models, and the video generation guide in OpenAI's docs, read on November 16, 2026, covers only the retired Sora 2 models.
Can Sora 2 generate real people?
It could not, and now it generates nothing. The API rejected real people, public figures and input images with human faces, and blocked human likeness in character uploads by default. Gemini Omni Flash 1.2 takes up to 3 character references per request, and its docs list no likeness block like Sora's.
The live replacements covered here are the SeeDance 2.0 API, the Veo 3.1 API and the Kling 3.0 API, each with a playground for testing a prompt before writing code.


