Sora Is Gone: OpenAI Switches Off the Sora API. Here Are the Video Models to Use Instead

TL;DR: OpenAI's Sora API stopped serving requests on 24 September 2026, five months after the Sora app and website closed on 26 April. Sora survives only as a world-model research programme. Anyone who built a product or a workflow on it now needs a replacement: Google's Veo 3.1 and Gemini Omni Flash, ByteDance's Seedance 2.5, Kuaishou's Kling 3.0 and Alibaba's Wan 3.0 are the candidates, and none of them is a drop-in swap.

When OpenAI announced the two-stage shutdown of Sora earlier this year, the reasoning was blunt. As The Decoder reported, the company is redirecting resources toward coding tools and enterprise customers, and toward "a super app rolling ChatGPT and other tools into one package." The consumer app closed on 26 April 2026. The API followed on 24 September.

For most people this is a footnote. For teams that built generation pipelines, social tools or agency workflows on Sora, it is an outage with a deadline that has already passed.

What happens to your Sora videos

According to the same report, OpenAI let users download their content and export videos and images directly from the Sora library while the services were live. The company said it had not decided whether there would be a final export window after the cutoff dates, and that after all deadlines pass, user data is permanently deleted. If you have not exported your library, check your OpenAI account now; do not assume a grace period exists.

Sora itself is not dead as a research direction. OpenAI says the work continues as a world-model programme aimed at eventually "automating the physical economy." That is a very different product from a text-to-video API, and there is no announced date for anything developers can call.

The replacement candidates

The video-generation market did not stand still while Sora was winding down. Here are the models that were shipping and documented as of late September 2026, with what each is good at.

Google Veo 3.1 and Gemini Omni Flash

Google now sells video generation two ways through the Gemini API. Veo 3.1 is the dedicated video model, billed per second of output: the Gemini API pricing page lists a range from $0.10 to $0.60 per second depending on the resolution and variant (the cheaper "Lite" tiers at the low end, 4K at the high end). Gemini Omni Flash is Google's multimodal model that can produce video as one of its outputs; it is billed per token, with the pricing page listing $1.50 per million input tokens and $17.50 per million tokens of video output. Google's changelog records an Omni 1.1 Flash release on 27 August 2026 that added video extension and interpolation and output up to 4K.

Best for: teams already on Google Cloud, anyone who wants a single API for text, image and video, and workloads where per-second billing is easier to forecast than token billing.

ByteDance Seedance 2.5

ByteDance's Seedance 2.5 announcement (31 July 2026) describes single-pass generation of clips up to 30 seconds with synchronised audio, dialogue lip-sync driven by quoted lines in the prompt, and up to 50 reference assets per generation (30 images, 10 videos, 10 audio clips). The API opened publicly on 7 August at 480p and 720p. It is the strongest option if your work depends on character or product consistency across shots, because the reference system is far more flexible than Sora's was.

Our prompt library has a large collection of Seedance prompts with previews, which is the fastest way to see how the model responds to cinematic direction, camera moves and stylised action.

Kuaishou Kling 3.0

Kuaishou's Kling 3.0 release adds native 4K output, clips up to 15 seconds, multi-shot generation inside a single prompt and lip-sync across five languages. Turbo and Omni variants followed in June. Kling has the most mature creator ecosystem outside China of the three Chinese models, and its multi-shot feature is the closest thing to Sora's storyboard-style prompting.

Alibaba Wan 3.0

Wan 3.0, announced on 24 August 2026 (coverage), generates 30-second single-shot videos and accepts up to 20 reference assets. Unlike earlier Wan releases it is API-only; Alibaba has not published weights. It is worth a look if your input is documents or slides rather than prose prompts, which is the use case Alibaba is leading with.

Midjourney V8.2

Midjourney remains an image-first tool with video features, and V8.2 has been the default model since 24 July 2026 according to Midjourney's documentation. There is still no public API, so it is not a Sora replacement for developers, but for individual creators who used Sora through the app it is the closest in spirit.

How to choose

Whichever you pick, budget time for prompt migration. Sora prompts leaned on natural-language scene description; Seedance and Kling reward explicit camera direction and reference assets; Veo's per-second pricing changes how you think about clip length. Run the same ten prompts through two candidates before committing, and keep the outputs so you can compare quality when the next model lands, which at the current pace will be within the quarter.

Sources

• The Decoder, OpenAI sets two-stage Sora shutdown
• Google, Gemini API pricing and Gemini API changelog
• ByteDance Seed, Introducing Seedance 2.5
• Kuaishou investor relations, Kling AI launches 3.0 model
• Midjourney documentation, Version