Sora is gone. Here's where AI video goes next.

Sora is gone. OpenAI closed the app on April 26, 2026 and switched off the API on September 24, without naming a replacement. If you made ads, shorts or pitch films with it, here is what happened, what to do with the clips you still have, and which tools are worth a look now.

Short version: export anything you still have, expect no successor from OpenAI, and pick a new tool by what you actually make: vertical ads with sound, long single takes, or work for clients that needs clean commercial terms.

What happened, and when

On March 24, the official Sora account announced on X that OpenAI was “saying goodbye to the Sora app.” The dates came later. OpenAI's help center lists April 26 for the app, and the API deprecations page lists September 24 for the Videos API and all five sora-2 model versions. The “recommended replacement” column on that page is empty.

One copy is still running for now. Microsoft's model retirement schedule lists its preview sora-2 deployment for retirement on October 15, 2026, also with no replacement.

Why OpenAI shut it down

OpenAI never published a full explanation. A spokesperson told CBS News the decision came down to focus and growing demand for computing power, and that the Sora research team would move to world-simulation work for robotics.

Press reports fill in the rest. Citing The Wall Street Journal, TechCrunch reported that Sora cost about $1 million a day to run, and that its users peaked near one million before falling below 500,000. The shutdown also ended Disney's December 2025 deal, a $1 billion investment plus licensing for more than 200 characters, before any money changed hands, according to TechCrunch and CBS. Sora 2 also drew copyright and deepfake complaints from its launch, but OpenAI didn't name them as the reason.

If you used Sora: five things to do now

• Export what you have. OpenAI's export tool lives at sora.chatgpt.com/sunset. OpenAI says it will email users before any final export window, and that Sora data is permanently deleted after that. No final deadline has been announced yet, so don't wait for one.
• Ask about refunds the normal way. There are no automatic refunds. OpenAI points users to the standard ChatGPT subscription refund process, and says leftover ChatGPT or Sora credits can be spent on Codex.
• Find every Sora call in your stack. Changing the model name won't fix a broken integration, because there is no successor model. Check your code, automations and any Azure deployments before October 15.
• Re-check rights before you reuse old clips. Clips built on someone's likeness or a licensed character need the same permissions they needed before. The tool shutting down doesn't change that.
• Keep the labels. Sora output carried visible watermarks and C2PA provenance data by default, per OpenAI. Whatever tool you switch to, keep labeling AI video as AI.

Where to go now

These are the tools a former Sora user is most likely to consider, with what each vendor publishes today. We aren't ranking them by quality. Prices are what each vendor lists as of October 13, 2026, and they change often.

• Google Gemini Omni Flash. On paper, the closest match to Sora's make-then-remix workflow: 3 to 10 second clips, 16:9 or 9:16, sound on by default, and conversational edits. In the API it costs about $0.10 per second at 720p, per Google's pricing page. Paid consumer plans start with Google AI Plus at $4.99 a month.
• Google Veo 3.1. 4, 6 or 8 second clips with native audio and 9:16 support. The Standard and Fast versions extend in 7-second steps. Its Lite version is the cheapest major API we found at $0.05 per second. Every Google clip carries an invisible SynthID watermark, per Google's docs.
• Kling 3.0. The current Kling model, with multi-shot generation. Kling 4.0 is announced for October 2026, but so far only a limited “Flash” early-access version is out, per Kling's blog and third-party trackers. Through fal, Kling 3.0 Pro costs $0.112 to $0.168 per second depending on audio.
• Runway. Its own Gen-4.5 model plus partner models (Seedance 2.5, Kling 3.0, Gemini Omni Flash) under one plan. Its terms, updated September 15, 2026, allow commercial use on every plan, free included. The catch: Runway may use your inputs and outputs to train its models.
• ByteDance Seedance 2.5. The long-take option: single clips of 4 to 30 seconds, 9:16, native audio, about $0.47 per second at 720p on fal. Some articles say it isn't available in the US, but Dreamina's official account announced a US launch in August, and it is also offered through fal and Runway.
• Luma Ray3.2. Built for finishing: 16-bit HDR, EXR export and up to 16 keyframes. Plans start at $30 a month, with commercial use included on every plan, per Luma's pricing. There is no free plan.
• MiniMax Hailuo (H3). Released July 31: up to 15 seconds at 2K with stereo sound, per MiniMax. Its open weights reportedly come with a license that excludes the US, so US teams should use the paid app or API.
• Pika 2.5. Known for effects and lip-sync. No watermark on any plan, but the commercial license starts at the Creator plan, $35 a month, per Pika's pricing.
• Alibaba Wan 2.2. The open-weight pick. It's Apache 2.0 and its 5B model runs on a single 24 GB GPU. The newer Wan 3.0 is API-only.

Commercial use, at a glance

This is what decides whether a clip can go in a paid ad. The last two rows come from third-party summaries, because the official terms pages couldn't be read. Check them yourself before relying on them.

One more thing to know before a client asks: in the US, purely AI-generated footage may not be protected by copyright at all, per the US Copyright Office's 2025 report on AI and copyrightability. A license to use a clip is something you can give. Full ownership may not be.

Quick picks, by job

• Vertical ads with sound: Gemini Omni Flash, Veo 3.1 or Seedance 2.5. All three make 9:16 natively with audio.
• One long, unbroken take: Seedance 2.5, at up to 30 seconds per clip.
• Client work with the clearest terms: Runway or Luma.
• Lowest API cost at volume: Veo 3.1 Lite, at $0.05 per second.
• Running it on your own hardware: Wan 2.2.

These picks are based on each vendor's published specs, not on side-by-side quality tests.

The lesson: don't build on one model

OpenAI opened Sora to ChatGPT subscribers in December 2024. Twenty-one months later, its API was gone. Any model can disappear on that kind of timeline, and the tools on this list will keep changing too.

The parts of a Sora project that survive the shutdown are the ones the model didn't make: the script, the reference images, the storyboard, the edit, the taste. Keep those in a form any tool can use, and switching models means re-rendering, not starting over. The model is the camera. The direction is yours.

Facts are from the linked sources, checked on October 13, 2026. Usage and cost figures are press reports, not OpenAI numbers. Rows marked “reported” come from third-party summaries. This isn't legal advice. We aren't affiliated with any company named here.