Sora is shutting down: the AI video tools to switch to before the September API cutoff

If you built anything on Sora, you have a hard deadline. OpenAI discontinued the Sora web app and mobile experience on April 26, 2026, and the Sora 2 API stops working on September 24, 2026 (OpenAI Help Center). There is no announced successor. When that date passes, every request your pipeline sends to the Sora endpoints returns an error, and OpenAI has said user-generated content will be deleted after the deadlines pass.

The reason matters for how you plan the next twelve months. OpenAI has cited the economics directly: Sora cost roughly $1 million per day to run against thin revenue, and coverage of the shutdown put total revenue near $2.1 million while usage dropped (the-decoder). A frontier video model got switched off because the unit economics did not work. That is worth remembering when you pick your replacement: bet on a vendor whose video product is core to the business, not a side experiment.

The window is now short. As of late August 2026 there are roughly four weeks left, so this post is a migration guide rather than a feature bake-off. If you want the full head-to-head on quality and pricing, read our four-way comparison of Sora, Veo, Runway, and Kling. Here the job is narrower: what replaces Sora for each kind of work, and how to move without breaking production. Versions below are a mid-2026 snapshot, so treat the exact numbers as a point in time and re-check before you sign anything.

The two decisions you actually have to make

There are only two: what is your deadline, and what did Sora do for you.

On the deadline, split your work into two buckets. Anything already running on Sora needs a migration finished and tested before September 24, 2026. Anything new should not start on Sora at all, even though the API still answers today. Building fresh on a tool with a known kill date is throwing away effort.

On the second question, be specific. "AI video" is not one job. Sora was strong at physics and motion realism, and people used it for very different outputs: single hero shots, storyboard-driven sequences, multi-language dialogue, high-resolution commercial work. No single replacement wins all of those, so the right move is to map each use case to the tool built for it.

Which tool replaces Sora for which job

Here is the practical mapping. Match your use case to the row, not the brand.

Physics-heavy hero shots -> Hailuo 2.3 (MiniMax). Sora's calling card was believable motion and physical interaction. Hailuo 2.3 is the closest match on that specific strength, which makes it the natural target for product-in-motion shots, collisions, fluids, and anything where the eye catches fake physics immediately. Reviewers also flag it as the strongest generation-capacity-per-dollar option among subscription tiers (diyai.io), which helps if physics shots mean lots of iteration.

Storyboard and character workflows -> Runway Gen-4.5. If your Sora work went through a shot list with recurring characters, Runway is the lowest-friction move. Gen-4.5 offers reference-driven character consistency, camera and motion control, and the Act One and storyboard tooling that maps cleanly onto a creative-suite process. Runway also has studio and platform partnerships (Lionsgate, AMC, Adobe) that make it an easier internal sell for teams that need an approved creative vendor.

Single 4K commercial and broadcast shots -> Veo 3.1 (Google). When the deliverable is a finished commercial asset, Veo 3.1 is the safest pick. It is the only model in the mainstream set producing native 4K, and its audio is the cleanest, generating 48kHz synchronized dialogue rather than sound effects alone (imagine.art). For broadcast and paid media where resolution and audio sync are non-negotiable, that combination is hard to beat. Veo 3.1 Lite and Fast tiers give you cheaper options for drafts and internal review.

Multi-language dialogue -> Kling 3.0 Omni (Kuaishou). If your pipeline produces the same scene in several languages, Kling 3.0 Omni is the only option here with a native multi-shot dialogue timeline and five-language lip-sync, and it tends to be the cheapest per clip. The caveat is procurement: Kling runs on Chinese infrastructure, which is a blocker for some regulated enterprises and major studios. If your legal and data-residency requirements rule that out, treat Veo as the fallback for multi-language work and dub separately.

A sixth name is worth watching but not committing to yet. Dreamina Seedance 2.0 (ByteDance) currently tops the Artificial Analysis text-to-video-with-audio arena, so it is rising fast, but it is not yet on many enterprise approved-vendor lists. Use it for experiments, not for a September deadline.

One more filter before you commit: procurement, not just quality. The output you can generate is only useful if your legal and security teams will approve the vendor. Kling and the ByteDance and Alibaba options run on Chinese infrastructure, which many regulated enterprises and major studios cannot use for data-residency or supply-chain reasons. Runway's studio partnerships (Lionsgate, AMC, Adobe) and Google's enterprise footing make Veo and Runway the easier internal sell in those environments. Sort the shortlist by what your organization can actually buy, then compare quality inside that filtered set. A model you cannot get through procurement is not a real option no matter how good the demos look.

If you need to keep it in-house

Every option above is a hosted API, which means the same category of risk that just bit Sora: someone else decides whether the service keeps running. If your reason for control is compliance, data residency, or simply not wanting a vendor to sunset your pipeline again, look at open-weight models you can run yourself. Wan 2.5 and 2.6 (Alibaba) ship open weights and can run on your own infrastructure. Quality sits below the hosted frontier, and you take on the GPU and MLOps cost, but nobody can send you a discontinuation notice. For a regulated team that got burned by Sora, that trade is often worth it. A common pattern is a hybrid: an open-weight model self-hosted for the sensitive or high-volume baseline work, and a hosted frontier model called through the same abstraction layer for the handful of hero shots where quality has to be the best available.

A migration plan that fits the time left

The original version of this plan assumed a quarter. It no longer fits, so here is the compressed sequence, written with dates rather than week numbers so you can see where you actually are.

By August 31, export and inventory. Pull everything out of Sora first. Export your generated content before the deletion deadlines using OpenAI's export flow. Then list every place Sora is called: which scripts, which product features, which scheduled jobs. You cannot migrate what you have not found.

By September 7, classify and choose. Tag each Sora use case against the mapping above: physics, storyboard, 4K commercial, or multi-language dialogue. Most teams find they were quietly using Sora for two or three different jobs, which means two or three different replacements rather than one. That is normal and it is cheaper than forcing everything onto one tool that only fits half your work.

By September 14, wrap it and run both in parallel. Do not wire your application directly to a second vendor's API. Put a thin internal interface in front of video generation so the rest of your code asks for "a clip" without knowing which model produced it. This is the single most valuable thing you take away from the Sora shutdown. The next time a provider changes terms or shuts down, you swap one adapter instead of rewriting your pipeline. It is a few days of work that pays for itself the first time a vendor surprises you.

Then route real jobs through the new tool while Sora still answers, and compare the output against your existing footage. This is where you catch the gaps: a prompt style that worked on Sora may need rewriting for Veo or Kling, and lip-sync or aspect-ratio settings rarely map one-to-one. Fix those while you still have a working reference.

By September 21, cut over and confirm. Flip the abstraction layer to the new default, watch a full cycle of production jobs, and only then decommission the Sora path. Leaving three days of buffer before the API stops is deliberate: it covers the problem you did not predict.

If you are reading this in mid-September, drop the parallel run. Export first, pick one replacement for your highest-volume job, wrap it, and move that path across. A rough cutover you control beats a clean plan that runs out of calendar.

A quick word on prompts, since this is where most of the hidden work lives. Each model was trained differently, so the phrasing that produced a clean result on Sora will not translate one-to-one. Veo responds well to explicit camera and lighting direction; Kling wants clear shot boundaries for its multi-shot timeline; Hailuo rewards physical detail in the description. Budget time to rebuild a small library of proven prompts for whichever tools you pick, and store those prompts alongside your abstraction layer so the knowledge does not live in one person's head. This is the part teams underestimate, and it is the reason the parallel run matters even when the calendar is tight.

One planning note for anyone budgeting this: a clean migration is mostly plumbing and prompt work, not model magic. The teams that struggle are the ones who hard-coded a single API deep into product logic. The teams that move in a week are the ones who wrapped it. If you are building new video features now, wrap them from day one.

What this means beyond Sora

The specific lesson is a deadline. The general lesson is that hosted generative video is still a young, consolidating market where even a well-funded frontier model can be switched off for economic reasons. Design for that. Keep an abstraction layer, keep a second vendor qualified, and for anything genuinely critical, keep an open-weight fallback you can host. Sora will not be the last model to go dark, and the teams that treat model providers as swappable will barely notice the next one.

Start with the export today. Everything else is easier once your content is safe and your use cases are mapped, and the export is the only step whose window closes for good.