Sora Has Shut Down: Where to Make AI Video Now

If you made video with Sora, the tool is gone. The consumer app closed in April 2026, and on 24 September 2026 OpenAI switched off the Sora 2 API as well. This post covers what closed and when, who is affected besides OpenAI's own users, and where each kind of video job should go next. Everything here is as of September 2026.

What happened, and when

OpenAI announced the end of Sora on 24 March 2026. The shutdown came in two stages.

• The Sora app and web experience closed on 26 April 2026, according to OpenAI's help-centre notice on the discontinuation. Video generation inside ChatGPT closed on the same day.
• The Sora 2 API closed on 24 September 2026. OpenAI's API deprecations page lists the Videos API and every Sora model ID (sora-2, sora-2-pro and their dated snapshots) with that shutdown date. In the column for a recommended replacement, it lists nothing.

That empty column matters. When OpenAI retires a text model, it usually names a successor. Here it named none, so OpenAI has no video model to move to. The reasons were widely covered in March. The Hollywood Reporter also reported that Disney left its OpenAI deal after the closure was announced. For creators the reason matters less than the result: there is no first-party OpenAI video product left.

Who this actually affects

Three groups feel it, and not all of them know yet.

Direct Sora users. If you made clips in the app, it has been closed for five months, and you have probably already moved. The practical question is whether you kept your source files. Any project built around Sora's look and characters now has to be recreated in another model. No other model reproduces Sora's style exactly.

Developers on the API. Any product, script or n8n flow that called the Videos API broke on 24 September. There is no drop-in endpoint with the same shape, so you have to rewrite the integration for another vendor.

Customers of apps that resold Sora. This group is the easiest to miss. Many multi-model creator apps offered Sora 2 alongside Veo and Kling, and they reached it through the API. When our researchers checked vendor pages on 25 September, a day after the shutdown, several still advertised Sora 2. That included Hailuo's plan cards, Higgsfield, PixVerse, HeyGen, invideo AI, Hedra, Krea's pricing page and Filmora. Some of those pages are simply out of date. Either way, you cannot generate Sora video in them now. If you chose a plan partly because it listed Sora, you are paying for a model you cannot use. Check what the plan still gives you, and ask the vendor about credits if Sora was the reason you bought it.

We have removed Sora from every app's model list on the model map, so it now shows only models you can use today.

How to choose a replacement

Don't look for "the new Sora". Sort your work by job and pick a model for each one. No model leads at everything, and the same model costs very different amounts in different apps. Three questions narrow the choice:

1. Do you need sound? Most 2026 flagship models now generate dialogue, effects and ambience with the picture. A few still don't.
2. How long does one shot need to be? Base clip lengths range from about 5 seconds to 30 seconds per generation.
3. Where will it be published, and under what licence? Free tiers are usually watermarked and often non-commercial. Check commercial use and watermarks before you pick a plan, not after.

Text-to-video with audio

This is where most Sora users were, and there are now several strong options.

Veo 3.1 🧠 ModelPaid

Google DeepMind's text- and image-to-video model with native sound, up to 4K

Google DeepMind · 2025-10

8.8 Visit ↗

Veo 3.1 (Google) generates picture and sound in one pass: speech, effects and ambience. It outputs 1080p or 4K in landscape or native vertical. Base clips are 8 seconds, and scene extension builds longer sequences. Veo is available in more places than almost any other model. Google's own options are the Gemini app, Google Flow and the Gemini API. It also appears in Adobe Firefly Boards, Higgsfield, Krea, Leonardo.Ai, Luma's app, invideo AI and many others. For developers, the Gemini API charges per second with audio included. At the time of writing, that is $0.40 at 720p or 1080p for Standard and much less for Fast and Lite. It is the most direct replacement for the Sora 2 API.

Kling 3.0 🧠 ModelFreemium

Kuaishou's video model with native multilingual audio, lip sync and 15-second clips

Kuaishou · 2026-02

8.6 Visit ↗

Kling 3.0 (Kuaishou) makes clips up to 15 seconds with multi-shot storyboarding. It has native audio, including dialogue in several languages with lip sync. It is strong at character motion and talking scenes. You can use it in Kuaishou's own Kling AI app, where paid plans remove the watermark and add commercial use, or inside Firefly, Runway, Higgsfield, Krea, Luma's app and ComfyUI partner nodes.

Seedance 2.5 🧠 ModelPaid

ByteDance's video model: 30-second clips with audio, many references and local edits

ByteDance Seed · 2026-07

8.6 Visit ↗

Seedance 2.5 (ByteDance) gives you the longest single takes: up to 30 seconds per generation. It generates audio and video together and accepts many reference images and clips in one request. You reach it through ByteDance's own apps (Dreamina, and CapCut for Seedance 2.0, though not in every country) or through partner apps such as Higgsfield, Krea, Runway, Luma's app, invideo AI and Pika. ByteDance's launch material says little about watermarks or likeness rules, so read the terms of the app you use.

Also worth testing: Runway Gen-4.5 (model page), which added native audio and multi-shot scenes of about a minute. MiniMax H3 (Hailuo), which makes 15-second clips with stereo sound at 2K. Vidu Q3 and PixVerse V6, both of which have native audio and suit stylised and short-form work.

Image-to-video and directed shots

If you animated stills or product shots in Sora, the job is to keep a start image faithful while it moves.

• Kling has start- and end-frame image-to-video and motion control. It is a safe default for animating a key visual.
• Luma Ray3.2 lets you place up to 16 keyframes inside one clip and renders native HDR. It suits exact timing for VFX plates or ads. It has no native audio, and Luma's app has no free plan.
• MiniMax H3 accepts up to 9 reference images plus video and audio clips. Seedance 2.5 takes up to 30 images. Both help when a product or character must stay the same from shot to shot.
• Veo 3.1 takes several reference images for character and product consistency.

The uncomfortable part: reference images reduce drift but do not remove it. Logos, small text and hands still slip. Budget for retries, and check every frame of a product shot before you publish.

Avatars and talking heads

Some creators used Sora for "person talks to camera" clips. A general video model is the wrong tool for that. Dedicated avatar apps give you consistent presenters, script-driven editing and a consent process.

• HeyGen and Synthesia are the presenter-style leaders. Both run consent checks on custom avatars and remove the watermark only on paid plans. See avatar videos for the full list.
• Hedra combines Character-3 avatars with a large multi-model studio.
• For a talking character inside a scene rather than a presenter, Kling 3.0 and Veo 3.1 both do dialogue with lip sync.

Whatever you use, a realistic synthetic person saying something a real person didn't say needs a label on YouTube, TikTok and Meta. The disclosure rules by platform cover the details.

Editing footage you already have

Sora was a generator, not an editor, but many 2026 models now edit existing clips. That changes what "making AI video" means.

• Runway Aleph does video-to-video editing inside Runway's editor, next to Gen-4.5.
• Kling 3.0 Omni edits existing footage and keeps elements consistent across shots. Its June 2026 upgrade added 4K editing.
• Seedance 2.5 can fix one detail at a given timestamp without regenerating the whole clip.
• MiniMax H3 and Wan 3.0 both accept plain-language instructions to change one element and leave the rest alone.

If you mainly cut real footage, start in your editor. Adobe Premiere now offers Firefly and partner models, including Veo, Kling, Runway and Luma. Descript and CapCut also have generation built in.

Open and local options

The Sora shutdown is a reminder that a hosted model can disappear. If that risk worries you, keep one open-weight model in your toolkit.

LTX-2.5 🧠 ModelOpen source

Lightricks' open-weight video model with synchronised audio and 4K output

Lightricks · 2026-08 · open weights

7.8 Visit ↗

• LTX-2.5 (Lightricks) generates synchronised audio and video. It runs on consumer GPUs with 16 GB of VRAM and supports multi-shot generation. The weights are free for organisations under $10M in annual revenue. Larger companies need a paid licence.
• Wan 2.1/2.2 (Alibaba) are Apache-2.0 open weights and the default open video models in ComfyUI. The newest Wan release, 3.0, is API-only.
• HunyuanVideo 1.5 (Tencent) is a smaller model that is easy to fine-tune. It has no native audio and uses a custom licence.

Open models still trail Veo and Kling on realism. But nobody can switch off a model you run yourself.

For developers: rebuilding a Sora 2 integration

If your product called sora-2 or sora-2-pro, you need a new endpoint and a new cost model. Three options have published per-second prices:

MiniMax, Kling and Alibaba also sell APIs, but we could not find a published per-second rate for MiniMax H3 or Wan 3.0. Wrap your integration so the model is a setting you can change without a rewrite. Don't let this migration lock you into one vendor. The MCP servers for fal, Replicate and Runway make it easier to call several models from an agent.

What we'd do this week

• List every place Sora appeared in your work: saved prompts, templates, n8n or Zapier flows, client style guides. Replace each reference with a named model.
• Test the same three prompts in Veo 3.1, Kling 3.0 and Seedance 2.5 inside one multi-model app. That shows you the difference in quality for your kind of content before you pay for three subscriptions.
• Check your plan. If you pay for a multi-model app that still advertises Sora 2, confirm what you can actually use and ask about credits.
• Re-read the licence of whichever app you move to. The model does not decide commercial use and watermarks; the app's plan does.
• Label realistic synthetic people and scenes on every platform that asks for it.
• Keep one open model (LTX or Wan) working in ComfyUI as a fallback.

Read the official docs for…

• The Sora API shutdown date and affected model IDs: OpenAI API deprecations
• Veo per-second API pricing: Gemini API pricing
• Runway plans and Gen-4.5 credit costs: Runway pricing
• Per-model credit costs for Ray, Veo, Kling, Seedance and MiniMax in one app: Luma pricing