Your AI Vendor Just Deleted an Endpoint. Build Like It Will Happen Again.
On Monday, Google switched off three Imagen 4 endpoints. Not deprecated, not discouraged. Off. Anything still calling imagen-4.0-generate-001 or its ultra and fast siblings stopped working on 17 August.
Six days from now, on 26 August, OpenAI removes the Assistants API and points people at the Responses and Conversations APIs instead. Sora 2 and the Videos API go on 24 September, with no replacement listed. A long list of older GPT snapshots follows on 23 October. The GPT-5 snapshots that everyone standardised on last August retire on 11 December. Google shut down the Gemini 2.0 Flash family back on 1 June.
That is five separate retirement dates inside seven months, from the two vendors most small companies build on.
This is the operating rhythm, not a bad quarter
It is tempting to read that list as vendors behaving badly. It is more useful to read it as the actual shape of the market. Model providers are shipping faster than they can carry old versions, and carrying old versions costs them GPU capacity they would rather sell. The notice periods are honest enough (OpenAI commits to at least six months for generally available models) but six months is short if nobody in your company is reading deprecation pages.
The mistake is not picking the wrong vendor. The mistake is treating the model layer as stable ground when it is the least stable layer you have. Your database will still be there in three years. Your accounting system will. The exact model string in your config file will not.
Where it actually hurts a smaller company
Big companies have someone whose job includes watching vendor changelogs. A 30 person company does not. So the failure mode is specific and predictable: the AI feature was built once, by one developer or one agency, six months ago. It works. Nobody touches it. Then it silently returns errors one Monday morning, and the person who built it has moved on or moved to another client.
The second cost is quieter. Migration is rarely a rename. Google’s replacement image models take different parameters and produce different output, so the prompts and the post-processing both need review. Moving off the Assistants API means rethinking how state and threads work, not swapping a URL. Budget days, not minutes.
Four things worth doing this month
Find every model string you depend on. Grep your repos and your automation tools for model names. Most companies are surprised by how many places one got hardcoded, including inside Zapier or Make scenarios that no developer owns.
Put them behind one config value. One place per system that names the model, environment variable or config table, never scattered through the code. This is a two hour job that makes every future migration a ten minute job.
Own your prompts and your evaluation cases. Keep twenty real inputs with the output you expect, in a file, in your repo. When you swap models, you run those twenty and see what changed. Without them, “does the new model still work?” is a feeling rather than an answer.
Put one calendar reminder in. Quarterly, someone opens the deprecation page for each vendor you use and checks the dates against your list. Fifteen minutes, four times a year.
The trade-off worth being honest about
None of this argues for abstraction layers or multi-vendor gateways. For most SMEs those cost more than they save and add a component you now maintain. Pick one vendor, go deep, and simply keep the seams where the swap happens visible and cheap.
The point is not to avoid depending on a vendor. You will depend on one. The point is that the dependency should take a day to move, not a quarter, and that somebody in your company should know the date before the vendor’s calendar tells you the hard way.
We have written before about why AI projects fail on delivery rather than technology, and about the integration work being the real project. This is the maintenance half of that same argument. If you want a look at where your current AI setup is quietly hardcoded, that is part of what our free assessment covers.