OpenAI shuts down Sora, retiring its video generation service
Sora 2, released on September 30, 2025, reached its end about a year later. On April 26, 2026, the Sora website and app went offline; on September 24, 2026, the Sora API and the Sora 2 models were formally shut down. The Sora line's standalone video generation business is thereby retired, roughly a year after it began.
First, what the thing is. A text-to-video model takes text — and sometimes images — as conditioning and returns a sequence that is coherent over time, and it is usually judged on a handful of axes: visual quality, controllability, and whether characters and scenes stay consistent as clips get longer. Inference for these models is also among the heaviest of any modality: a few seconds of video means denoising repeatedly across both space and time, costing far more compute and memory than generating the same amount of text. That shapes the business model too — if it cannot get cheap, it is hard to spread through large-scale free use.
Timeline: from launch to shutdown
There was advance warning. On March 24, 2026, OpenAI notified developers that the Videos API, along with the Sora 2 model aliases and snapshots, were being deprecated and removed from the API on September 24. That is exactly six months between notice and shutdown, with the consumer-facing website and app going offline earlier, on April 26. The list is specific: the Videos API itself, plus the aliases and snapshots sora-2, sora-2-pro, sora-2-2025-10-06, sora-2-2025-12-08 and sora-2-pro-2025-10-06. What stands out is the "recommended replacement" column in that same table: it is empty — just a dash where other rows name a successor model.
What does that mean in practice? OpenAI's deprecation documentation names no replacement video model for this shutdown. Applications that called Sora 2 through the API therefore have no officially designated successor; developers still relying on it have to look elsewhere for video generation rather than simply swapping a model name and carrying on.
How the shutdown process works
The same documentation lays out OpenAI's general rules for retiring models. Unless safety or compliance concerns require something faster, generally available models get at least six months' notice; specialized variants — chat variants, Codex variants, deep research variants — get at least three months; and preview models, identified by "preview" in the name, may be retired with as little as two weeks, which is why the company advises against using them for business-critical production workloads. Notification comes by email and through the docs page, with blog posts for larger changes.
Against the Sora timeline: notice on March 24, execution on September 24 — exactly the six-month minimum that applies to generally available models. That also helps explain why the consumer website and app went offline on April 26 while the API side waited until September: the two follow different clocks, and developers get the longer migration window.
For engineering teams, a six-month notice period is not a courtesy. The docs spell out what it is for: evaluating recommended replacement models, testing application behaviour and completing migrations before a model becomes unavailable. This case is different, though — the replacement column is empty, so what actually had to happen over those months was finding an external option rather than upgrading to a newer in-house model. That makes the evaluation heavier: prompt conventions, parameter sets, output formats, moderation and billing all differ from provider to provider.
There is an engineering detail in the docs as well: by calling the list models endpoint (GET /v1/models) you can read each model's optional shutdown_date field and find out in advance whether a given model is slated to go away. The field holds a date string in YYYY-MM-DD form, or null when no shutdown date has been announced. For teams with model names hard-coded in configuration, wiring that field into monitoring is far calmer than waiting for the email. OpenAI also notes that in some cases developers may be able to provision dedicated capacity for continued access after a model's shutdown date.
Worth noting, the same page queues up a batch of contemporaneous deprecations: older GPT Image models shut down on December 1, 2026, certain GPT-4 series models on January 14, 2027, and text-to-speech and transcription models on January 6 and February 26, 2027 respectively. The reminder is that Sora's shutdown is not an isolated event but one entry in a continuously running retirement process — the pace of model turnover has itself become part of the cost of using a platform.
The other side of the lane
Veo 3.1's reach also says something in itself: video generation is turning from a standalone product into a feature embedded across many products. It appears simultaneously in a chat assistant, a filmmaking tool, a short-video platform, an office suite and a cloud platform, so users may not know which model they are using but keep using the capability. For that kind of distribution, model replaceability is a design premise rather than an afterthought — swapping in a stronger version should be imperceptible on the user side. On the other side of the same lane, Google's Veo 3.1 is still expanding. A January 13, 2026 update added native audio, 9:16 vertical framing, improved image-to-video and 1080p and 4K upscaling, across Gemini, Flow, YouTube Shorts, Google Vids, Vertex AI and the Gemini API. On specs, it supports 4-, 6- and 8-second clips, video extension, first and last frame conditioning, and up to three reference images.
Put the two side by side and you get this: one company has closed both its consumer video generation product and its API, while the other keeps pushing into more surfaces. The facts are what matter here — attributing the outcome to each company's own product decisions is safer than guessing at motives; the service shutting down, the empty replacement column and the competitor still shipping are all publicly checkable.
One more thing is easy to underestimate: video generation APIs are not interchangeable. Another vendor having a similar endpoint does not mean swapping the model name will do — the shape of the conditioning input, the clip length and resolution tiers, and how results are retrieved can all differ, and any link in the pipeline that assumes "this works like what I had" will surface during migration. The time-saver is to stand up a low-traffic side path during the notice window and surface the differences early.
From a product standpoint, this kind of capability has always had two commercial forms: an editor-style product for ordinary users and an API for developers. The first lives or dies on user scale and retention, the second on call volume and unit cost — and video generation is among the most compute-intensive modalities there is. Whether either is sustained long-term often does not move in step, and Sora's arc over the year demonstrates that asynchrony: opening with the consumer side first, then closing out on the API side.
The practical takeaway for anyone building a product is plain: design as if models will disappear. Do not assume an endpoint will always be there, keep a switchable alternative for critical paths, and wire deprecation notices and signals like the shutdown_date field into monitoring. If a product leans heavily on video generation, then when choosing a provider, treat "is there a replacement model" and "how long is the notice period" as being as important as price and quality. The rest is a matter of iteration.