On 24 September 2026, OpenAI switches off the Sora API for good. Not because it broke. Not because nobody used it. It simply reached the end of its published lifespan, and every business workflow wired into it stopped working that day.
This is the part of AI that almost nobody explains to business owners: the AI tools you buy come with expiry dates, and those dates are usually published months in advance in documents you have never read.
What is AI model deprecation?
AI model deprecation is the process of an AI provider formally retiring a model, so that it stops being available for new work and is eventually switched off entirely. The model does not degrade or break. The company simply stops running it, on a date it announces in advance, and anything built on that model must move to a replacement.
Think of it the way you think of a mobile network shutting down 3G. Your old handset still switches on. It just cannot connect to anything any more, because the thing on the other end no longer exists.
Most providers use a sequence of labels before the switch-off. Anthropic, for example, moves a model through four stages: active, legacy, deprecated, then retired. Each stage is a warning. "Deprecated" means it still works but is on the way out. "Retired" means it is gone.
Why do AI companies switch off models that still work?
Running an old AI model is expensive. Every retired model frees up the specialised chips, engineering attention and safety monitoring that a provider would rather spend on its newest model. Older models also carry outdated safety behaviour and weaker security handling, which becomes a liability the longer they stay live.
There is a straightforward commercial logic underneath. A provider that keeps twelve model versions alive has to test, patch and defend twelve of them. A provider that keeps three alive can charge less and move faster.
This is also why newer models usually arrive cheaper per unit of work. The industry has been pushing traffic onto smaller, faster tiers. If the difference between model tiers is new to you, our guide to what a "Flash" AI model is covers how the cheap tiers work.
What actually happened when OpenAI retired Sora?
OpenAI announced the Sora shutdown on 24 March 2026 and executed it in two stages: the Sora web and app experiences closed on 26 April 2026, and the API followed on 24 September 2026. That gave roughly six months of notice from announcement to final switch-off.
The detail that matters for a business is what happened to the material. According to OpenAI's own help documentation, after the discontinuation and after a final export window closes, data associated with a user's use of Sora is permanently deleted. Unused Sora credits were redirected to another OpenAI product rather than refunded.
So the shutdown was not just "a tool disappeared". It was a tool disappearing, a file library disappearing behind it, and a prepaid balance turning into store credit for something else.
A Hong Kong marketing agency that had built a client video pipeline on that API had three jobs to do inside that window: find a replacement, rebuild the workflow, and export every asset before the door shut. Any one of those left undone became a client problem in October.
Which AI shutdowns are already scheduled?
Several widely used models have published retirement dates in the second half of 2026. These are not rumours. They sit in provider documentation, and they apply to any tool or automation that calls those models by name.
Retirement dates already announced
--- 26 August 2026: OpenAI's Assistants API retired, replaced by the Responses API.
--- 24 September 2026: the Sora API discontinued, ending programmatic access to Sora 2 and Sora 2 Pro.
--- 23 October 2026: GPT-4, GPT-3.5 Turbo, o1 and o1-pro shut down.
--- 11 December 2026: the August 2025 snapshots of GPT-5 and o3 retire.
--- Earlier in 2026: the GPT-4o family was removed from ChatGPT on 13 February, the API alias was disabled on 17 February, and final enterprise access ended on 3 April.
The October date is the one that deserves attention. GPT-3.5 Turbo and GPT-4 were the default choices for a very large number of systems built between 2023 and 2025. A chatbot, a quotation generator or a document summariser built in that period may still be pointing at those exact model names today.
How much warning will you actually get?
Notice periods vary by provider and by how mature the model is. As a working rule, generally available models get months of warning and preview models get weeks. The warning is real, but it is delivered to developers, not to business owners.
Typical notice by provider
--- OpenAI: around six months for generally available models, roughly three months for specialised variants, and about two weeks for preview models.
--- Anthropic: a commitment of at least 60 days' notice for publicly released models, plus a published "not sooner than" date for each active model. Opus 4.8, for instance, carries a date of no earlier than 28 May 2027.
--- Across the industry: 30 to 90 days is a common notice window.
Here is the catch. That notice arrives by email to the developer account holder, or as a line in a changelog page. If your AI tool was set up by a vendor, a freelancer or a staff member who has since left, the warning lands in an inbox nobody is reading. The first sign you get is a customer telling you the chatbot has stopped replying.
What should a Hong Kong business owner do about it?
You do not need technical skill to protect yourself. You need a short written record of what your business runs on and who gets told when it changes. Most of the damage from a model retirement comes from nobody knowing the dependency existed.
Five practical steps, none of which require writing code:
--- List what you actually use. Write down every AI tool in the business, what it does, and who set it up. Include tools a department bought on a credit card without telling anyone.
--- Ask one question of each vendor. "Which AI model does this run on, and what happens to my service when that model is retired?" A vendor who cannot answer that has not planned for it either.
--- Make sure the notice reaches a live inbox. Provider warnings should go to an address the business controls, not to a departed employee or an agency's generic account.
--- Know where your data lives and how to get it out. The Sora case shows that an export window can close permanently. Test the export before you need it.
--- Prefer vendors who handle the migration for you. If a supplier's contract says they keep the service working across model changes, the expiry date becomes their operational problem, not yours.
The context here matters for Hong Kong specifically. A Dah Sing Bank survey of more than 340 local SMEs, conducted in May 2026, found that roughly 23% had already adopted AI or generative AI, with marketing and content creation at 56% of use cases and customer service at 42%. Those are exactly the workflows that sit on named models. Our breakdown of how Hong Kong SMEs are really using AI has the fuller picture.
Common questions about AI model deprecation
The most frequent misunderstanding is that deprecation only affects developers. It affects anyone whose business process depends on a specific model, whether or not they know the model's name.
Does this affect me if I only use ChatGPT in a browser?
Partly. Consumer apps usually migrate you to the newest model automatically, so nothing breaks. What changes is behaviour: the new model may answer differently, follow your saved instructions differently, or produce a different writing style. The February 2026 removal of the GPT-4o family from ChatGPT is the clearest example of users noticing a personality change overnight.
Will my automation just switch to a newer model by itself?
Not usually. Automations that name a specific model keep calling that name until it returns an error. Some platforms auto-upgrade; many do not. This is worth confirming in writing rather than assuming.
Is a retired model gone forever?
For practical purposes, yes. Once a hosted model is retired, the provider stops serving it and there is no way to bring it back on your side.
Should this stop me adopting AI?
No. Every business tool has a lifecycle, including the accounting software and the phone system. The difference with AI is the speed: lifecycles are measured in months rather than a decade. Planning for replacement from day one is the whole adjustment.
The takeaway
AI model deprecation is not a technical curiosity. It is a scheduled interruption to your business, published in advance, in a place you are unlikely to be looking.
The businesses that get hurt are not the ones using old models. They are the ones who never wrote down which models they were using. A single page listing your AI tools, the models behind them, and the inbox that receives the warnings converts a future emergency into a calendar entry.
Technology that expires does not have to feel cold. Someone should be reading the changelogs so you do not have to. We understand AI. UD stands with you.
Find out what your business is running on
If you are not sure which AI tools your business depends on or what happens when one of them retires, start with a quick assessment. We will walk you through it step by step, from mapping what you already use to planning what replaces it.
Reviewed by the UD AI team.
