How to avoid AI vendor lock-in

On 14 October, OpenAI told developers that GPT-4o and GPT-4.1 will stop working in its API on 14 January 2027 (OpenAI deprecations (opens in new tab)). That's three months. The same page promises at least six months' notice for generally available models, unless safety or compliance needs a faster timeline.

If you've built something on one of those models, a customer service flow, a document pipeline, a weekly report, you've now got a quarter to move it. Christmas included.

This isn't a post about OpenAI behaving badly. Every major provider retires models, and three months is plenty for a team that planned for it. It's about the teams that didn't plan for it, which is most of them.

In June we wrote about why the single-vendor default stopped holding. This is the practical side: what it costs you, and how to set things up so it doesn't.

You're renting it, and that changes the maths

ChatGPT, Claude, Gemini and the rest share one trait that's easy to forget once they're working well. You don't own any of it. You're renting access. The provider decides how it behaves, what it costs, and when your version stops existing.

That's fine on its own. Renting is how most businesses use software. The trouble is the assumption that comes with it: that what works today is what you'll be working with next year, so it's safe to build your processes around one provider without asking what a change would cost.

The tools aren't the problem. They're genuinely good. For work that needs real judgement, like reading between the lines of a complex document or drafting something where quality decides the outcome, they do in seconds what used to take a skilled person hours. And they're easy to start with. No technical team, no setup, up and running in an afternoon.

That ease is exactly why the setup decisions get skipped.

Four costs that only show up after you've committed

Your version gets retired on their timetable

Models aren't permanent products. Providers retire older versions on their own schedule. GPT-4o came out of ChatGPT in February 2026 (OpenAI (opens in new tab)), and the API version follows in January, on the three months' notice above.

When the version your process depends on goes, it isn't a software update you click through. Someone has to check whether the replacement behaves the same way, rewrite the instructions that don't carry over, and fix whatever broke. If AI runs through several parts of the business, that happens every time a retirement lands. The provider gives you the date. Sorting out your own setup is your job.

It can change how it behaves without a release note

This is the one that catches people out. In April 2025, OpenAI shipped an update to GPT-4o in ChatGPT that made it agree with and flatter users far too readily, and rolled it back four days later (TechCrunch (opens in new tab)). Through August and early September 2025, three infrastructure bugs intermittently degraded Claude's answers, and Anthropic's own postmortem (opens in new tab) says the early reports were hard to tell apart from normal variation.

Neither was a new version, and nobody had to opt in. In practice, a workflow your team spent weeks tuning can start producing something different. The proposal template drifts. The way incoming enquiries get sorted shifts. You find out when something looks wrong, or when a process downstream starts failing, and by then it may have been running that way for days.

If you build directly on a provider's API, you can pin a fixed version. That protects you from model updates, though not from bugs like Anthropic's, and only until that version is retired. The chat apps your team uses every day don't give you that option.

Every question your team asks leaves the building

When someone types into one of these tools, the text goes to the provider's servers, gets processed and comes back. For most questions, that's fine. Now think about what your team actually asks AI to help with: client proposals, financials, HR issues, customer data, contracts.

Business plans from the major providers don't train on your data by default, and they come with a data processing agreement. Personal accounts are a different story. Several consumer plans use conversations for training unless someone switches it off, so if your team is using their own logins for work, that's worth knowing today. And under UK and EU GDPR, sending personal data to an AI provider makes them your processor. "I assume it's fine" isn't an answer you can give if anyone asks.

It's cheap in testing and isn't at scale

Most of these tools charge by usage. In a trial, the bill is close to nothing. Once AI is processing documents, answering queries and running background tasks all day, the same pricing turns into a real monthly line, and it's rarely modelled before the setup is chosen.

The price isn't yours to set either. In November 2025, Anthropic cut the price of its top model by two-thirds (Anthropic (opens in new tab)). Good news, that time. But a provider that can cut prices by two-thirds can also raise them, and a budget built on today's rate is a forecast on a number someone else controls.

The businesses caught by these aren't careless. They made sensible choices early on, without asking what happens when those choices need to change.

Most of your AI use doesn't need the premium tier

None of this is a case for stopping. It's a case for being deliberate about which jobs need which tool.

Some work clearly earns the premium tier: pulling conflicting information together, writing where quality decides the outcome, anything that doesn't fit a template. There, the best tool pays for itself.

Most day-to-day AI use isn't that. Summarising documents and meeting notes. Sorting routine customer queries. Drafting from a clear brief. Pulling fields out of standard forms. The smaller, cheaper models from the same providers handle these at a fraction of the price. Sending everything through the most expensive option isn't thoroughness, it's paying for capability you aren't using.

How to set it up so you can move

None of this needs a technical team or an overhaul. It needs a few different decisions at setup, the same way you'd think about any key supplier. You might have a preferred one. You'd still want to be able to switch if their prices doubled or they stopped making what you need, without rebuilding the operation.

Match the tool to the job

List what your team actually uses AI for. It usually falls into a handful of buckets: reading and summarising, drafting, answering routine questions, organising information. Put the premium tools on the jobs that need them and the cheaper ones on the rest. That's not cutting corners.

Make switching a decision, not a project

When AI goes into a process, decide at the start what it would take to swap provider. Done well, that's a configuration change. Done badly, it's months. The point isn't to switch often. It's that when a price jumps, a version gets retired on three months' notice, or something better turns up, moving is an option you actually have.

Write instructions about the outcome, not the tool

Instructions that describe what you want produced, rather than working around one tool's quirks, carry across to another tool mostly intact. And keep a handful of real examples of good output for each process. When a version changes or behaviour drifts, you can check the new output against them in an afternoon, instead of finding out from a customer.

The question worth asking first

The useful question isn't "which AI tool should we use?" It's three smaller ones. What do we actually need AI to do? Which of those jobs genuinely need a premium tool? And if our provider gave us three months' notice tomorrow, what would we have to rebuild?

Those aren't technical questions. They're business planning ones, and the third is the one most businesses can't answer yet.