Home › Blog

Plain-English Briefing

Three AI services switch off this month — and nobody emailed you

By Charlie Essex · 8 August 2026 · 4 minute read

Between now and the end of August, three AI services stop working. A much bigger group follows in October. If anybody built you a tool in the last two years, checking which one it runs on is a twenty-minute job worth doing this week.

This one flatters my argument, so let me put the awkward part near the top: I sell a box that carries on working when a supplier changes its mind. There is a section further down about where that does not help, and it is the one I would read first if I were you.

What actually happened

Claude 3 Haiku on Google Cloud stops on 23 August. OpenAI’s Assistants API stops on 26 August. Moonshot retires its older Kimi models on 31 August. In October, GPT-4o, GPT-4 and most of the older OpenAI range go with them.

None of those names will mean much to you, and that is precisely the problem. They are engines. What you bought was a chatbot, or a quoting tool, or something that reads documents and gives you the gist — sold under a friendlier name, by somebody who chose the engine on your behalf and had no particular reason to tell you which one.

Why this keeps happening

Model lifespans have gone from roughly two years to somewhere between six and twelve months. One recent OpenAI model lasted five. The labs are competing on capability and it costs them real money to keep an old model running for the minority still on it, so the old one goes.

Retirement is also not the same as nobody using it. Months after GPT-4o was pulled from the consumer ChatGPT app, security researchers still found it running inside a large share of business cloud setups — more than any newer model from the same company. The date something is announced dead and the date it stops holding up somebody’s Tuesday morning are rarely the same date.

Three things this means for you

Ask one question this week. If a developer or an agency built you anything that talks to an AI, ask them which model it calls and what happens when that model retires. The answer may well be “nothing, we moved you months ago”. Ask anyway, and get it in writing, because the notice period is running whether or not anybody has told you.

A working tool can stop working with nobody at fault. Nothing is hacked, no invoice is missed, no one does anything wrong. The engine is simply switched off at the far end, and the first you know about it is a thing that used to work not working.

The cheap tiers go first. Older and lower-cost models are retired before flagships, because they are the ones the vendor is least paid to keep alive. If a tool was chosen on price, it is nearer the front of the queue than you would like.

Where owning does not help

A box in your building does not freeze time. Open models age too. Better ones arrive, the one you installed gradually stops being the sensible choice, and somebody has to decide when to move — and once you own it, that somebody is you. Anyone selling you “install it and forget it for a decade” is selling you something.

So the difference is narrower than it suits me to claim. It is not that nothing changes. It is that nothing stops on a morning you did not choose, and no notice period runs out in an inbox you were never on.

And if your team uses AI lightly — a few people, a few times a week, nothing going in that you would mind a stranger reading — then a subscription and a reminder in the diary is a perfectly good answer, and this month changes nothing about that. We have worked the comparison out properly, including an honest section on where renting wins.

Where this came from

• OpenAI — model and endpoint deprecations, in their own words
• Moonshot AI — the Kimi model list and platform sunset notice
• Anthropic — model deprecations and retirement dates

More plain-English briefings on the Inferred blog. If you would rather talk it through, the fifteen-minute call is free and I will tell you if there is nothing worth doing.