Sora in short: what happened to it
Sora is closed for good. OpenAI announced the end of the project on 24 March 2026; the web version and the app stopped working on 26 April, and the API on 24 September 2026. OpenAI released no replacement model of its own, so the question of what to use instead of Sora can only be answered by third-party generators.
If you still have videos in your account, OpenAI points to the export page at sora.chatgpt.com/sunset; after the export window the data is deleted permanently (OpenAI help page "What to know about the Sora discontinuation").
This piece is not news about the shutdown but a working reference: which models as of October 2026 fit the jobs sora-2 was picked for.
Which models replace Sora as of October 2026
There is no single replacement: different generators suit volume, on-screen text, long scenes and polishing footage you already have. Prices are official API rates in USD; resellers charge differently.
| Job | Model | What matters |
|---|---|---|
| many short vertical videos | Kling 3.0 | clips up to 15 s, multi-shot; $0.08-0.17 per second depending on resolution and audio |
| signage, packaging, on-screen text | Veo 3.1 Fast | native 9:16 and audio; $0.10 at 720p and $0.12 at 1080p per second |
| a long scene in one pass | Seedance 2.5 | up to 30 s per generation, up to 50 references; via Runway API $0.30 per second at 720p |
| a scene from a brief, a deck or a website | Wan 3.0 | up to 30 s, takes documents and links; $0.10 at 720p and $0.20 at 1080p |
| cheap motion drafts | MiniMax H3 | up to 15 s with audio; $0.08 per second at 768p |
| polishing footage you already have | Runway Gen-4.5 and Aleph | masks, repainting a finished clip; Gen-4.5 about $0.12 per second |
What launched in recent weeks
The market did not freeze after Sora: two notable models came out over September and October that are worth keeping on the test list.
- Kling 4.0: Kuaishou opened early access on 28 September 2026, with a full launch promised in October; native 30 seconds, 10-bit HDR and multiple keyframes are claimed. For now it is subscription access, not a stable API for batches.
- Vidu Q4 Preview from Shengshu: released on 7 October 2026, priced from $0.05 per second at 540p up to $0.40 at 4K, with up to 15 references per generation.
- Gemini Omni Flash: since May 2026 this is the model generating video in the consumer Gemini app, while Google's API still offers Veo 3.1.
How to pick a replacement in one evening
The choice is made on your own prompts, not on reviews. Five checks give an answer faster than any ranking.
- One object across three consecutive scenes: measure the share of frames where the product or the face does not drift. Models with references (Seedance 2.5, Wan 3.0) hold up better here.
- On-screen text: a label or a sign has to stay readable in the background too. Veo 3.1 is still ahead of most models on this point.
- Audio: speech with lip sync and background in one pass. Not every model supports Russian in native audio, so voiceover is sometimes still done with a separate synthesis.
- The price of a usable frame, not of an attempt: if a good result takes five generations, the real cost per second is five times the price list.
- Length: 15 seconds in one clip, or three 5-second pieces stitched together. Every join is extra editing work.
How to move a pipeline off sora-2
What moves is the script, not the prompt string: the same wording on another model gives a different frame, so prompts are rewritten for the specific generator.
- Put your prompts and storyboards into one table: model, length, resolution, result. With it, migration takes days; without it, weeks.
- Run 20 working prompts on two candidates and compare how many attempts each one needs for a usable frame.
- Recalculate the budget at the selected model's price per second, adding a third on top for rejects.
- Check the licence of the plan: on free tiers some services ban commercial use and leave a watermark.
- Pick a backup model: a generator is a rental, and any of them can shut down, as Sora itself showed.
What a batch costs on a replacement
Example: 50 vertical videos of 10 seconds each in 1080p with audio, generation only, no rejects.
Kling 3.0 with audio: 500 seconds × $0.17 ≈ $85. Veo 3.1 Fast: 500 × $0.12 ≈ $60, but 8-second clips have to be stitched. Wan 3.0: 500 × $0.20 ≈ $100.
Rejects usually take a quarter to a third of the generations, and Russian voiceover, subtitles and editing for three platforms cost more than the generation itself. So the price per second is an input to the quote, not the price of a video.
What next
Detailed parameters for each model are in the 2026 guide to AI video models. If you need a ready stream of videos rather than a choice of generator, look at the AI video production for business service and the vertical video format: handcrafted videos turnkey start at $190.