The AI feature you depend on is somebody else's contract

You did not choose most of your AI stack. You chose one feature that worked, and everything under it came along. Siri taps into ChatGPT for what Apple calls “more in-depth answers for certain requests” [3]. Your writing tool summarizes with a model its maker does not own. Your automation platform calls an API you never signed anything for. Each of those is not a product you bought. It is a commercial arrangement between two companies, and your working day is standing on top of it.

That arrangement can end without anyone doing anything wrong. Deals expire, get repriced, or become awkward when the two firms start selling into the same market. This guide is for the solo operator, freelancer or small-team owner on self-serve plans, where nobody negotiated a notice period on your behalf. If you have an enterprise agreement with a named account manager and contractual change windows, most of this is already handled in your paperwork and you can skip it. Everyone else is relying on a policy page they have never read.

Three layers, three different amounts of warning

Every AI capability you use sits on one of three layers, and the layer decides how much warning you get before it changes.

The top layer is a consumer app feature: a mode, a button, a project sidebar inside ChatGPT or Claude. There is no published notice period for these. They appear and disappear with releases, and the only signal you get is the interface changing under you.

The middle layer is a bundled integration, where one company ships a feature that runs on another company’s model. Apple’s ChatGPT extension is the clearest example. Apple’s own documentation describes it as something you switch on: “If the ChatGPT extension is off, you can set it up in System Settings” [3]. It also tells you that “You don’t need a ChatGPT account, but if you have one—free or paid—you can connect to your account” [3], and that when you use it, “Your request and attachments, as well as limited data associated with the request, such as current time zone, country, device type, language, and feature being used when making the request, are sent to ChatGPT” [3]. Read that carefully. The control is a toggle in Apple’s settings, the model is OpenAI’s, and neither Apple’s page nor OpenAI’s published deprecation policy says anything about how long the arrangement lasts [1][3].

The bottom layer is a direct API call, and it is the only one with a written promise. OpenAI’s deprecation policy gives at least 6 months for generally available models and at least 3 months for specialized variants, while noting that “Preview models, identified by preview in the model name, may be retired with much shorter notice, such as 2 weeks” [1]. Anthropic states that it notifies customers “providing at least 60 days’ notice before model retirement for publicly released models” [2].

Those policies are real and they get used. Anthropic deprecated claude-opus-4-1-20250805 on 5 June 2026 and retired it on 5 August 2026 [2]. OpenAI lists 23 October 2026 as the shutdown date for gpt-3.5-turbo-0125, gpt-4-turbo, o1-2024-12-17 and o3-mini-2025-01-31 [1]. Entire surfaces go the same way: OpenAI notified developers of the Assistants API deprecation on 26 August 2025 and removes it a year later, on 26 August 2026, pointing users at the Responses and Conversations APIs instead [1].

The uncomfortable part is the inversion. The layer that gives you the most warning is the one you had to do the most work to build on. The layer that gives you none is the one that felt like it came free with the device.

A bundled integration is a deal, not a product

The Apple and OpenAI relationship is the case study, and it is worth being precise about it rather than dramatic. On 11 July 2026 Apple filed suit in the US District Court for the Northern District of California against OpenAI, its hardware subsidiary io Products, OpenAI chief hardware officer Tang Yew Tan and engineer Chang Liu, alleging a “coordinated pattern of misconduct at an institutional level” involving confidential product designs, manufacturing methods and supplier information. The complaint brings four counts under the federal Defend Trade Secrets Act and two breach-of-contract claims, and Apple is asking for preliminary and permanent injunctions, preservation of electronic evidence, return of its information, damages and a jury trial [8].

Now hold that next to the other fact: Apple still documents how to turn on the ChatGPT extension in macOS [3]. Two companies can be in federal court against each other while the integration between them keeps working normally. That is the actual lesson, and it is more useful than any prediction about who wins. Nothing in your daily experience of the feature tells you anything about the health of the deal underneath it.

You do not need a lawsuit for this to matter. The general pattern is that a bundled feature carries the union of both companies’ strategic changes: either party’s repricing, either party’s product pivot, either party’s decision that the other is now a competitor rather than a supplier. Your exposure is the sum, and your notice period is whatever the more abrupt party feels like giving.

What locks you in is the material you never exported

If you asked most operators what would make switching AI vendors hard, they would say the model. It is almost never the model. It is the accumulated material: the prompts you refined over months, the reference documents you uploaded into a project, the instructions you wrote once and forgot, the chat history you search when you cannot remember how you solved something.

Both major vendors will give that material back to you, and the mechanics are worth knowing before you need them. In ChatGPT, you go to Settings, then Data controls, then “Under Export data, select Export” and confirm [4]. What arrives is a file that “includes your chat history and other relevant account data” [4]. Two details matter more than the steps: “Exports can take up to 7 days to arrive” and “The download link expires 24 hours after you receive it” [4]. In Claude, you click your initials, open Settings, go to the Privacy section and click “Export data”; the help page says “Data exports include conversation data and the user data for your account” and that “The download link will expire 24 hours after delivery” [5].

What neither export page documents is a way in. Both describe getting your data out; neither describes handing another vendor’s file to the product and getting your setup back [4][5]. The closest thing that exists is manual and explicitly hedged: Anthropic documents transferring memory by asking your other assistant to “Output everything in a single code block so I can easily copy it” and pasting the result into Claude, while warning that “Memory imports are experimental and still in active development, and at this stage, Claude may not always successfully incorporate imported memories” [6].

That asymmetry should change where you keep things. A prompt saved as a text file in a folder you own moves to a new vendor in a copy and a paste. The same prompt living only in a vendor’s project sidebar moves via an export request, a 7-day wait, a 24-hour download window and a manual rebuild. The words are identical. The switching cost is not.

Price the move before you are forced to make it

When people compare vendors they compare rate cards, which is the wrong number to focus on. Look at the spread inside a single vendor’s own lineup: GPT-5.6 Sol is $5.00 per million input tokens and $30.00 per million output, while GPT-5.6 Luna is $0.20 and $1.20 [7]. That is a 25x gap between two models from the same company. If token price were the binding constraint, you would already be tuning it.

The number that actually decides whether you can move is your own time. Count the workflows that touch a specific vendor: the automations with an API key in them, the documents whose instructions assume one model’s habits, the client deliverables produced through a particular tool. Estimate honestly how long each takes to rebuild and verify somewhere else, then multiply by what an hour of your time is worth. That figure is your real lock-in, and unlike the token price, nobody publishes it for you.

If that number is small, stop worrying and get back to work. If it is large enough to be a bad month, the fix is not to abandon the vendor. The fix is to reduce the number of hours, which usually means pulling prompts and reference material out of the vendor’s interface and into files, and putting the model name in one place instead of 12.

Build for 2 weeks of notice

Assume the worst published number applies to you, because on the bundled layer there is no published number at all. Two weeks is a reasonable planning figure, since that is what OpenAI reserves for preview models [1] and it is more than a partnership announcement will give you.

Keep the model identifier in one configuration line rather than scattered through scripts, automations and saved prompts, so a swap is one edit and not an afternoon of search-and-replace. Keep every prompt and instruction set as a file in your own storage, treating the vendor’s copy as a cache. Take an export from each assistant you actually rely on now, while nothing is on fire, and download it inside the 24-hour window [4][5]. Once a quarter, run your 3 most important prompts against a second vendor’s model and read the output properly, so you find out where they diverge on a quiet Tuesday rather than during a client deadline.

None of this is a hedge against the specific Apple and OpenAI dispute, and it should not be sold to you as one. It is the ordinary cost of building on infrastructure that other people own, which is all of it.

What still goes wrong

Exports are less of a safety net than they look. An OpenAI export “cannot restore chats that have already been deleted” [4], and the combination of up to 7 days to arrive and a link that dies after 24 hours [4] means an export you request on the day bad news breaks is not an export you have. The only export that helps is the one you took months ago and can still find.

The second vendor is not a drop-in replacement, and no page will tell you how far off it is. A prompt tuned against one model produces different output on another, and the difference tends to show up in the places you check least. Anthropic’s own guidance on carrying memory across providers concedes the point directly, calling the process experimental and warning that imported memories may not always be incorporated successfully [6]. Budget rework time, not just migration time.

There is also a real cost to over-preparing. If you are one person, spending a week building an abstraction layer across three vendors to hedge a risk that may never fire is worse than spending that week on billable work. The proportionate version is small: files instead of sidebars, one config line instead of 12, one export a quarter. The published deprecation policies cover API models and API surfaces, not consumer app features or cross-company partnerships [1][2], so for anything on the bundled layer, the honest position is that you are exposed and the mitigation is keeping your material portable, not predicting the outcome.