October 23, 2026 was the date OpenAI set for the model name gpt-4 to stop answering. The notice went out on April 22, and like every notice on that page it carried a small second column. On the left sat the thing that was dying. On the right sat gpt-5.6-sol, the thing OpenAI recommended you use next. Procurement teams do not read that right-hand column as a strategy document. They should, because the best public evidence says that for most software, that column ends up being the whole decision.
Most advice about avoiding AI vendor lock-in concentrates on two moments, signing the contract and drawing the architecture diagram. A third moment matters more and nobody budgets for it. A vendor tells you your model has an end date, and for a few weeks a closed decision is open again. A retirement notice is a free rebid, handed to you by the incumbent. Almost everybody wastes it, and the thing that wastes it is the calendar.
Four in five came after the door shut
On September 25, Hyungjin Lukas Kim of Myongji University posted a preprint called When the Model Retires. He mined GitHub for commits moving off retired OpenAI, Anthropic and Google models and found 22,555 of them in 17,703 non-fork repositories. Of those, 5,139 matched an official retirement event.
His estimate is that 82% of genuine migrations were committed after shutdown. The median migration landed 39 days after the shutdown and 158 days after the announcement. His own sentence is better than any paraphrase: "four in five genuine migrations away from retired models happened after the model had stopped answering."
Then he looked at where they went. In his words, "only 8% of migrations switch provider." Between 68% and 74% stayed with the provider that had just retired their model, typically moving to the designated successor. Another 16% added a second provider alongside the first. The vendor pulled the model, sent a letter naming the replacement, and most of the software took the replacement.
The paper also checked whether pain teaches. Repositories hit by two or more retirements showed the same reactive share, 73%, at the first event and at later ones. They did move sooner after each shutdown, so they got faster at cleaning up. They did not get earlier.
The limits are real and you should hold them in mind. This is one author, an observational design, and public GitHub repositories, with no discussion of enterprise systems. A bank with a change board behaves differently from a weekend project. So look at what executives say about themselves. Zapier polled 500 U.S. enterprise executives and 89% said they could move off their primary AI vendor within four weeks. Two-thirds, 66%, had already tried a migration between AI platforms. Of those, 42% reported a smooth transition, and 58% said it failed outright or took significantly more effort than expected. Zapier sells integration software and has a stake in this story. Even so, the distance between 89% and 42% is the distance between a belief and a test.
Fourteen lines of code
The next number in the preprint should change what you buy. The median migration was 14 added lines and 6 deleted lines across 2 files. "Model identifiers are hard-coded in 94% of migrating applications," Kim writes, and only about 3% of migrations happened in code that already used an abstraction layer, while 4% introduced one.
The standard answer to lock-in is a gateway, a router, some neutral layer between your code and the provider. Kim himself notes that moving the model name into an environment variable or a registry would turn most migrations into a one-line change. He is right, and you should do it this week. Now run the arithmetic the other way. The swap was already close to a one-line change for the median project, and it still happened late and still went to the heir. Typing the new name was never the expensive part.
The expensive part is knowing the new model works. About 8% of migrations modified prompts. Only 4.5% of commit messages mentioned an evaluation or regression suite. Commit messages surely undercount testing, since nobody writes everything down. Silent failures are harder to wave away. In about 8% of genuine migrations, 14 of the 180 that human coders read, an exception handler had turned the retired-model error into a fallback response, such as a null result or a constant answer. The software kept running and said nothing useful, and nobody knew.
On this evidence, lock-in lives in the missing exam. If you have no test that says what good looks like for your claims process or your contract review, then every alternative model is unpriced. You cannot compare it to anything. The heir feels safe because it has the same API key, the same invoice and the same security review already on file, and because a familiar name stands in for evidence when you have none. That last sentence is my inference, and I would defend it in front of any board.
Some workloads have a second lock on top of the first. In Kim's data, prompt-only applications changed a median of 6 lines, while fine-tuned applications changed 693. OpenAI's page says fine-tuned versions follow the same schedule as their base models, and that on January 6, 2027, "Active existing customers will no longer be able to create new fine-tuning jobs on this date." If you fine-tuned on a vendor's base model, you own a thing that dies on someone else's date and cannot be remade the same way.
The shoppers and the heirs
There is an obvious objection, and Ramp supplies it. On October 6, Ramp's Ryan Stevens reported that the monthly share of firms switching providers, averaged over three months, reached a record 8% in September. He ties the latest shift to OpenAI's price cuts and new model launches starting in June 2026, and counts at least nine model launch announcements from Anthropic and OpenAI alone since then. He suggests the competition may be an early sign that models are becoming commodities. If firms move this freely, lock-in looks like yesterday's worry.
The two 8% figures measure different things and should not be compared directly. Ramp counts firms paying a provider, and the page does not say how many of those firms replaced a vendor and how many added one. Kim counts commits at the moment of forced migration. Read together, they suggest a pattern. Companies move when a better price or a better model shows up and they choose the moment. Software stays put when a deadline chooses the moment.
Plenty of finance teams will look at their Ramp-style spend data, see two vendors on the card, and conclude they have avoided lock-in. Two invoices tell you who you pay. They tell you nothing about whether the claims workflow on the second vendor passes the same exam as the first, or whether that exam exists.
The next ninety days are already printed
Vendors publish their calendars, which makes this easy to check for yourself. Look at what is on them now.
Anthropic notified developers on September 30 that claude-sonnet-4-5-20250929 retires on November 30, 2026, with claude-sonnet-5-5 as the recommended replacement. Anthropic's policy promises at least 60 days of notice for publicly released models. That date is just over three weeks from today.
OpenAI's page is longer. Older GPT Image models go on December 1. Older GPT-5 and o3 snapshots go on December 11. On October 14 it told developers using GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini that the models would be removed from the API on January 14, 2027, and it named gpt-6-luna as the replacement for all of them. GPT-5.3-Codex, GPT-5.1 and GPT-5.4-Nano follow on April 1, 2027.
That October 14 notice deserves a second look. OpenAI's stated policy gives generally available models "At least 6 months" of notice unless safety or compliance concerns require faster action. October 14 to January 14 is about three months, and the page gives no reason for the difference. I am not accusing anyone of anything. The published minimum and the published date simply sit on one page and do not obviously agree, and your contract team should ask which of them binds you.
Length of warning matters more than almost anything else in Kim's data. The reactive share ran 89% for Anthropic, 66% for OpenAI and 64% for Google after reweighting. The OpenAI Assistants API, which got a one-year notice, came in at 13%. His regression finds that each e-fold increase in notice length cuts the odds of post-shutdown migration by about three quarters. He calls notice length "the single policy lever most clearly associated with whether applications break." The caveat is that the event-level correlation, across only 13 events, misses significance at the 0.05 level (p = 0.06), so treat the direction as firmer than the size.
Now put that next to your own process. Suppose a notice gives you 60 days, building a test suite takes 30 days and security review of a second vendor takes 45, which puts 75 days of work on a 60-day clock before anyone has compared a single answer. The heir wins that race by default. Nobody chose it. Nobody on your team even got to make the case against it.
Vendors retire models for ordinary reasons. Anthropic says plainly that it retires models "to ensure capacity for new model releases," and it lists the downsides on the same page. I would rather have the capacity and the honest table than a museum of old weights. The notice is fair. Your failure to be ready for it is a choice you made somewhere upstream.
Your exam is on the shutdown list too
The most interesting lines on OpenAI's page have nothing to do with models. A June 3 notice covers three tools. Agent Builder, the Evals platform and the v1/prompts API with its reusable prompt objects are all scheduled to shut down on November 30, 2026, the same day Sonnet 4.5 retires on the Anthropic side. On the Evals platform, "Existing evals become read-only." on October 31, which was eight days ago.
Read that again as a lock-in story. The exam you would use to decide whether to leave a vendor was stored inside the vendor. It froze before the shutdown date of models it might have helped you compare. OpenAI's migration path points to Promptfoo, a testing framework that began as open source and that OpenAI announced it would acquire on March 9, 2026, with plans to fold the technology into its Frontier platform. Promptfoo may be an excellent tool. You should still keep your test cases in a format you could hand to something else, because the party that grades your exit should not also be the party you are exiting.
The guidance for reusable prompts is more candid than most consulting decks: "To migrate, move reusable prompt content into your application code." Agent Builder users are pointed to the Agents SDK or ChatGPT Workspace Agents, both of them OpenAI products. A lot of companies spent the past year putting their business logic in vendor-hosted consoles because the console was right there. Some of them learned this month that the console had an end date.
The request itself drifts as well. Anthropic's model deprecation page records that temperature, top_p and top_k were removed in API version 2026-10-21, where they return a 400 error, with prompting offered as the replacement. Your code that tunes those settings has to change even if you never leave Anthropic. Staying put has a migration bill. Since you are paying it anyway, you might as well get a rebid in return.
The law will remove a fee
Europe has an answer to lock-in with a date attached. Under the EU Data Act, Latham & Watkins wrote, "all switching charges will be prohibited beginning 12 January 2027." That is two days before OpenAI's January 14 removal date. The same analysis describes a customer right to start switching on a maximum of two months' notice, followed by a transition of a maximum of 30 days, and calls the definition of data processing services broad. Latham's piece dates from August 2025 and predates AI products as they exist today, so ask counsel to read the current text against your specific services before you rely on it.
Take the right if you have it. Then look at what it covers. A switching charge is a line on an invoice, and a regulator can delete it. No regulator can write your test cases, or tell you whether a new model reads your insurance forms the way your underwriters need. In the Zapier poll, only 30% said clearer pricing, features and contract terms would most reduce lock-in, and 26% said easier data transfers would. Those are real requests. They also describe the parts of the problem a contract or a statute can reach. The rest sits in your own building.
Avoiding AI vendor lock-in is a standing drill
Everything above points to one conclusion. AI vendor lock-in is what happens when your readiness runs slower than your vendor's calendar, and you fix it by changing your own timing. These are the moves I would make, in order.
Hold a bid on every notice
Treat each retirement notice as a procurement event that opens automatically. Three candidates go in, namely the named heir, a model from a rival, and an open-weight option where the workload allows it. The heir may well win. It should win on your exam, in writing, with a date. Say the rule is ten working days from notice to a signed test plan, and that a named executive owns the rule. You will be surprised how often the heir wins anyway, and how much better your renewal conversation goes when the vendor knows there was a bake-off.
Keep the exam in your own building
Your evaluation set is the asset. Real inputs, real expected outcomes and the failure cases your staff remember with a wince belong in your own repository, in a plain format, owned by the business unit that suffers when the answers go wrong. Prompts belong next to them. OpenAI just told its own customers to do the second part. Anything hosted by a model vendor should be treated as a convenient copy. Check this in the most boring way possible, by asking your team to export everything and run it against a different provider before the end of the quarter.
Buy notice like you buy uptime
Retirement policy is a commercial term, and the route you buy through changes it. Anthropic's page says partner-operated platforms such as Amazon Bedrock set their own retirement schedules, so status and dates can differ. Amazon's lifecycle page for models launched on or after September 7, 2026 lists two Legacy periods, 6 months and 45 days, and tells you to read the model card first. Going through AWS does not always buy you AWS's clock, though. The same page says "GPT-6.1 Sol follows the OpenAI first-party model's lifecycle and end-of-life terms." It also carries a trap for anyone keeping a dormant fallback, since in the Legacy period "existing customers may lose access after 15 days of inactivity." A backup you never call can vanish for that very reason.
Extensions exist, and they are sold. OpenAI says that "In some cases, developers may be able to provision dedicated capacity for continued access after a model's shutdown date," and tells you to contact sales. AWS mentions a "private arrangement between you and the provider." So when you sign, negotiate a contractual minimum notice per critical model, and a priced extension right, while you still have leverage. The price of an extension in week one of a negotiation and the price on the morning of a shutdown will not be the same number.
Make the running system hear it
Kim's suggestion to providers is that a warning should appear in API responses and SDK logs, so the notice "would reach the running system rather than an inbox." Do not wait for them. Assign every model name in production to an owner in a registry, with its retirement date beside it, and pull the dates from every route you use. Anthropic's Console will export usage by API key and by model as a CSV, which gives you a place to start. Then follow his blunter advice that "Exception handlers around model calls should fail loudly." An outage gets fixed by lunch. A constant answer can run for a month.
Pinning is worth a word too, because many teams treat it as a safety measure. "Pinning a dated snapshot without a retirement watch converts a reproducibility practice into an outage," Kim writes. Keep the pin and add the watch.
Rehearse one real exit a year
Pick one workload that matters, and move it to the second vendor for real, on a schedule nobody forced on you. Measure the days, the failures and the changes to prompts. Backups taught this lesson decades ago: a restore you never ran is a hope with a storage bill. The number from that rehearsal replaces the 89% confidence in your own leadership team with a fact, and it gives your CFO a price for the option you have been claiming to hold.
Architect the exit before the letter arrives
You cannot buy your way out of this with a tool. A gateway makes the swap one line, and the swap was already small. A second contract gives you a second invoice. A statute removes a fee. None of them can tell you which model should run your business on the morning after the notice, and that judgment is the only thing that turns a retirement into leverage.
So build it now. Decide who owns the model registry and who owns the exam. Write the rebid rule into your operating calendar. Take notice periods into your next renewal as a priced term and run the first rehearsal before the January dates arrive. These are architecture decisions, and they cross engineering, procurement, legal and the business units that own the outcomes, which is why they tend to fall between all of them until a deadline makes them for you.
Agor AI Advisory does this work with leadership teams, from the exam design to the contract language to the first live exit, and we have no model to sell you at the end of it. Your vendor has already written the name of your next model in a table. Decide now that you will be the one who reads it, tests it and has the standing to say no.
Sources
- When the Model Retires: An Empirical Study of LLM Migration in Open-Source Applications, Hyungjin Lukas Kim, arXiv, September 25, 2026
- Deprecations, OpenAI API documentation, notices through November 5, 2026
- Model deprecations, Anthropic Claude documentation, including the September 30, 2026 notice
- Model lifecycle, Amazon Bedrock User Guide, for models launched on or after September 7, 2026
- Businesses are changing AI model providers more often, Ryan Stevens, Ramp, October 6, 2026
- AI vendor loss would disrupt 3 in 4 enterprises, Zapier, 2026
- OpenAI to acquire Promptfoo, SiliconANGLE, March 9, 2026
- EU Data Act: Significant New Switching Requirements, Latham & Watkins, August 19, 2025