By: TitApeX BoHan /October 12, 2026

The AI Model in Your Software Has an Expiry Date: Settle AI Model Retirement Before You Commission Custom Software

The system you paid for is not broken, yet one morning it may simply stop working because the AI model behind it was switched off by its provider. This is not hypothetical. OpenAI shut down its Assistants API on August 26 and its Sora video API on September 24. On October 23, a batch of older models including gpt-4, gpt-4-turbo and o3-mini will stop serving requests. Anthropic notified developers on September 30 that Claude Sonnet 4.5 will retire on November 30. AI model retirement is now a crowded calendar. For a business commissioning custom software, the question is not whether it will happen, but who gets the notice, who does the replacement, and who pays. Here are four things to settle before you sign.

AI Model Retirement Is Scheduled, Not Accidental

A developer reads an AI provider's deprecation notice while a wall calendar marks several shutdown dates

An AI model differs from a traditional software component in one important way: it is a service with an end date. You do not own it. You rent it, and the provider decides when to take it back.

The two major providers state their rules openly:

OpenAI: at least 6 months' notice for generally available models, 3 months for specialized variants, and about 2 weeks for preview models. After shutdown, the model or endpoint is no longer accessible.
Anthropic: at least 60 days' notice before retirement for publicly released models. After retirement, requests to that model fail. When the model is used through Amazon Bedrock or Google Cloud, those platforms set their own schedules, so dates can differ.

It is not only old models that disappear. The Sora video API went from announcement to shutdown in six months, and the official page lists no replacement. We covered what that meant for video workflows in what Sora's shutdown teaches brands. The same logic applies to your order system, your support bot and your internal tools.

Why Swapping a Model Is Harder Than It Sounds

An engineer frowns at two screens where an old and a new AI model answer the same prompt in different formats

Many people assume changing models means editing one line of configuration. In practice, a new model may respond differently to the same instructions. The tone shifts, the output format changes, and a judgment that used to be stable starts to drift. Sometimes the parameters themselves change. Anthropic's documentation notes that newer models no longer accept adjustments to settings such as temperature and return an error if they are set.

A Zapier survey published in April shows the gap. It polled 542 U.S. enterprise executives and found:

89% believe they could switch AI vendors within four weeks.
About two-thirds have actually tried a migration, and 58% of them say it failed or took far more effort than expected.
Only 6% say they could stop using their main AI vendor without any disruption.

Two caveats: Zapier sells an automation platform, and the respondents were mostly large organizations. But the gap between "we think it's quick" and "it was hard" is likely to be wider at smaller companies, because most have no one watching provider announcements.

Four Things to Put in Writing About AI Model Retirement

A business owner and a developer discuss a project timeline on a laptop with a future model replacement date highlighted

1. Which models, on which platform

Ask the developer to list the full name and version of every model the system uses, and whether it is bought directly from the provider or through a cloud platform. Ask for the currently published retirement date of each. You do not need to understand the technical details. You need the list to exist and someone to keep it current.

2. Who receives the retirement notice

Providers send notices to the account holder. If the account belongs to the development firm and the engagement ended six months ago, that email may go unread. The safer setup is for your company to own and pay for the AI provider account, with the developer added as a member.

3. Who pays for the swap, and how it is priced

Changing a model involves at least three pieces of work: adjusting instructions, retesting, and monitoring after launch. State in the contract whether this is covered by the maintenance agreement or quoted separately, and if separately, at roughly what scale. This is the same point we made in the missing line on the quote: costs that arrive after launch should be discussed before launch.

4. Whether the system is built to be swapped

Ask the developer two questions. First, are all AI calls routed through one layer, so that changing a model means changing one place? Second, is there a saved set of real cases that can be run on the old and new model side by side? A practice shared recently in the developer community is to keep 30 to 50 real cases. With both in place, a swap takes days. Without them, it can turn into a small rebuild.

AI Compares, People Decide What "Just as Good" Means

A shop owner points at the AI reply samples that best match the brand tone while a developer looks on

During a model swap, the division of work is clear:

AI handles: running the test cases through old and new models, flagging where results differ, and watching for a sudden rise in errors after launch.
People handle: judging which differences are acceptable and which affect customers, deciding when to switch, and owning the outcome.

A new model's answers may be more fluent without being closer to your brand voice or your business rules. Only someone who knows the business can say what "just as good" means.

Conclusion: Write Replacement Day Into Today's Contract

A business owner reviews a maintenance plan with a developer on a tablet in a bright office

AI models will keep getting better, and they will be replaced faster. That is good news for businesses, as long as both the system and the contract are ready for it. A good custom software agreement names the launch date, and it also says who does what on the day a model retires. That is what TitApeX means by "AI handles perfection, people handle what moves us": technology turns over, so responsibility and judgment need to be arranged first.

If you are planning a system with AI features, or want to check how long the models in your current system have left, see our services or get in touch.