Sora is shutting down: API pricing, shutdown date, and where creators go next

Sora is not just leaving the consumer app. Its API has a fixed shutdown date, so production users need to move prompts, reference assets, and cost assumptions before the endpoint disappears. OpenAI has not named a one-for-one successor, and this article does not speculate about why Sora was retired.

The Sora shutdown in three steps

The source of record is OpenAI's Sora discontinuation notice. It directs users to export at sora.chatgpt.com/sunset as soon as possible rather than wait for a possible final export window.

What the final Sora API costs

The invoice is not the same as the cost of a usable shot. For iteration work, budget with output seconds x attempts x output rate, then add any input charges and finishing steps. The same retry problem appears across credit plans, which is why cost per usable result matters more than cost per generation.

For Scopeful's method for handling changing vendor pages, regional prices, and missing public rates, see how every price is verified.

Where Sora users can go

There is no honest single winner here. Each option below solves a different part of the Sora workflow, and none has been assigned a quality verdict without one controlled test set.

MiniMax H3: the clearest priced API candidate

MiniMax says H3 accepts text, images, video, and audio as context, generates native stereo sound, and produces clips up to 15 seconds at 2K. Its official launch page also describes text-to-video, image-to-video, multi-shot work, reference-driven editing, and video-to-video motion transfer.

That breadth makes H3 the most direct model to test first when a Sora workflow needs generation rather than editing alone. It does not make H3 a proven replacement. MiniMax's own examples and comparisons are vendor evidence, so character consistency, prompt adherence, motion, and usable-output rate still need an independent production test.

MiniMax called H3 an open model at launch but said the weights would follow. Treat its public API as the verifiable production path until the exact weight files, license, and hardware requirements for the intended deployment have been checked.

LTX-2.5: the ComfyUI and rented-GPU route

LTX-2.5 has no universal per-second price because the official path includes downloadable weights and local pipelines. The Lightricks repository supports text-to-video, image-to-video, keyframe interpolation, video transformation, audio-to-video, retakes, training, and ComfyUI. Its quick-start download is roughly 66 GiB across the model components, with quantization and CPU or disk offload available when memory is tight.

This route makes the most sense when a repeatable ComfyUI graph will render enough clips to justify instance startup, downloads, storage, and maintenance. At very low monthly volume, those operational costs can matter more than the lack of a model API fee. If managing the GPU becomes the job, the trade-offs in fal.ai versus Replicate are a useful fallback.

LTX-2.5 is open weights, not unrestricted public-domain software. Review the official LTX-2 license, including its commercial licensing threshold, before a company deployment.

FLUX 3: a watchlist, not a migration budget

Black Forest Labs says FLUX 3 Video is in early access. The announced plan covers video and audio generation and editing, keyframe-to-video, multilingual dialogue, longer multi-shot sequences, and eventual private and open-weight access. BFL's published 720p tests are explicitly preliminary, and there is no public API price.

That is enough to justify a test when access arrives. It is not enough to plan a September migration around it. A production replacement needs a stable access path, limits, terms, and a price before a launch demo can become a budget.

Wan 3.0 belongs on the rumor watchlist, not in the priced replacement table. As of August 12, Alibaba's official model release page still lists Wan 2.7 as the current commercial line, while the official Wan repository remains on Wan2.2. Until Alibaba publishes a Wan 3.0 model page, price, or weights, third-party launch claims are not enough for a production plan.

Grok Imagine Video 1.5: useful when the first frame already exists

Grok Imagine Video 1.5 is generally available, but its API currently supports image-to-video rather than text-to-video. That makes it relevant to product shots, concept frames, and still-image animation. It does not cover a Sora pipeline that begins with text alone.

The resolution ladder is public, but input is billed separately. Test total request cost and failed-shot rate with the same reference image before comparing it with an output-only headline rate.

Gemini Omni Flash and Kling 3.0 Omni are better treated as known baselines than new releases to chase. If either one already behaves predictably in a real workflow, keep it in the control group while testing H3 or LTX-2.5. Omni remains a public-preview, conversational editing option at roughly $0.10 per output second, but its current API limits still exclude audio references, video extension, and reliable processing of video references longer than 3 seconds.

Hosted API or self-hosted weights?

The choice is less about ideology than volume. Hosted APIs reduce setup and idle cost. Self-hosting can become economical when one stable workflow produces enough output to spread the setup cost across many renders.

Decision matrix

If Seedance is already in the pipeline, the Seedance 2.0 cost breakdown and unlimited-plan guide are better baselines than switching only because a newer model exists.

A migration plan before the API closes

1. Export Sora work now. Save finished outputs, source images, prompts, aspect ratios, duration choices, and any notes needed to reproduce the shot.
2. Choose one real workload. Do not test with launch-demo prompts. Use a job that includes the people, products, camera movement, text, and continuity problems that normally cause retries.
3. Pick two candidates from the matrix. One should match the input mode. The other should offer a genuinely different cost or control model.
4. Run the same test set. Record output seconds purchased, input charges, failed requests, render time, and how many outputs are usable without repair.
5. Compare cost per usable second. A cheaper quoted rate loses if it needs substantially more attempts or a separate repair pass.
6. Dual-run before September 24. Keep the Sora path available until the replacement passes the workload and export checks.

What Reddit reaction can and cannot prove

Early Reddit threads are useful as a list of problems to test, such as identity drift, weak physics, missed prompts, audio errors, or difficult local setup. They are still anecdotal, self-selected, and often compare different resolutions, quantizations, interfaces, and prompts.

In the recent threads I checked, H3 had the warmest overall tone, although users still reported wide-shot face problems, unwanted dialogue, and demanding hardware. LTX-2.5 was sharply divided between people who saw too little progress over 2.3 and people who valued its speed, audio, and more practical 1080p output.

The roughest repeated reaction was around the Grok Imagine Image 2.0 consumer update: users reported plastic-looking skin, generic faces, weaker reference handling, and anatomy regressions in one large complaint thread and a separate SFW workflow discussion. A few commenters preferred Fast mode, so this is not a unanimous verdict. xAI's public API docs also do not expose a separate Image 2.0 model ID, which is why this article calls it a consumer update rather than a verified API release.

That is why this article does not name a quality winner or call a model a flop. Community reaction can shape a test set. It cannot replace one controlled comparison on the same production brief.

Reader poll

If you have actually used one of these releases, add the failure you saw to the signal. Email links only preselect an answer; the vote is recorded after you confirm it here.