What happens when an AI model is retired
A model you use is being retired. Here is what changes in the app, what happens to conversations you already started, and what to check next.
A model retirement runs to a published schedule, and nearly all of that schedule is visible inside the app. It opens with a small amber pill reading "Deprecated" on the model's card, and once a shutdown date is set, it closes on a date that pill will tell you. In between, nothing you have already written stops working.
Model makers retire models. A version gets superseded and a shutdown date goes on the record. MultiChats carries 59 models from 17 providers, so somewhere in that list this is happening several times a year. Below is what it looks like from your side, in the order it actually happens.
The retirement, stage by stage
A warning appears, and it names names. The model keeps working and stays in the picker, but its card grows an amber "Deprecated" pill. Hover that pill on the web and the tooltip gives you two facts: the date the model will be discontinued, when one has been set, and the model we recommend in its place. One catch worth knowing: on a model your plan does not include, that hover shows the upgrade prompt instead. On phones the row carries the same amber badge and sinks to the bottom of the list, under everything still current. In the regenerate and branch menus on the web, retiring models are pulled out of their provider group and collected under a heading of their own.
The model details repeat the warning, more briefly. Open the small information button on a model card and the details panel carries a yellow block reading "Deprecated" and "May be removed soon". That block does not repeat the date or name the successor, so the tooltip on the pill stays the more useful of the two.
Choosing it hands you the replacement. From the moment a model is marked for retirement, selecting it resolves to its named successor. You can watch it happen. Tap the retired model in the picker and the selection highlight lands on the replacement's card, while the button you pick models with (in the composer on the web, in the header on mobile) reads the replacement's name.
On the shutdown date it leaves the list. Once the discontinuation date has passed, the entry is filtered out of the model picker, the regenerate menu and the branch menu, on web and on mobile. It stays in our registry after that, and that is precisely what keeps your old conversations pointing at something real.
Worth knowing: a deprecation warning is not a countdown to a broken chat. Nothing in your history stops working on the shutdown date, and nothing needs to be exported first.
A worked example: GPT-5
GPT-5 sits in this state right now. Its entry is marked deprecated, GPT-5.4 is named as the replacement, and the discontinuation date on its entry is 11 December 2026. Because that date is still ahead of us, GPT-5 is still in the picker with the amber pill on it, and choosing it already routes your request to GPT-5.4. Once the date passes, the GPT-5 card stops appearing anywhere you choose a model. Several other entries carry the same warning, the newest from late September 2026: Claude Opus 5 and Claude Sonnet 5 now hand you Claude Opus 5.5 and Claude Sonnet 5.5, and GPT-6 Sol and GPT-5.6 Sol both hand you GPT-6.1 Sol. None of those four has a shutdown date on its entry yet, so the tooltip names the replacement without a date, which is why the pill is worth reading rather than dismissing.
What happens to the conversations you already have
This is the part people actually worry about, and the short answer is that nothing is deleted, moved or rewritten. The longer answer is worth having, because one thing does change and we do not announce it.
The thread stays exactly where it is. It is not archived and not flagged. It opens and searches the way it did last week.
Old answers keep the model that wrote them. Every assistant reply is labelled with the model that produced it, and that label is stored on the message rather than derived from the thread. Scroll back through an old conversation and you still see GPT-5 next to the replies GPT-5 wrote. Retirement does not rewrite that history.
The next message goes to the replacement. The thread's stored model is resolved to its successor before the request is sent, and resolved again on our server before any model is called. You keep chatting in the same thread, with the same history behind you, answered by the successor.
What visibly changes is one line of text. The name on that model button. Open an old thread that was pinned to a retired model and that button reads the replacement instead of what you remember choosing.
On the web, your default quietly follows. Opening such a thread also updates your saved model preference to the replacement, so the next new chat you start begins there rather than reviving a retired name.
We do not announce the switch. There is no toast, no banner inside the thread, no email. So if a long-running conversation starts to feel different, in tone or in how it structures an answer, compare the name on the model button against the label on your older replies. A mismatch there is the whole explanation.
Four things to check when your model goes
Read the tooltip on the web before anything else. It is the only place where the shutdown date and the recommended replacement appear together. Everything else in the interface tells you half of that, and the phone picker never shows the date at all.
Compare the capability lists. The model details card lists what a model can take and do: vision, PDF, video, web search, reasoning, image generation and voice. Open it for the replacement and check that the things you actually used are still on the list. On phones the picker rows show the same icons apart from voice, so voice is worth confirming on the web card.
Assume nothing about a prompt you tuned. A prompt shaped over weeks against one model can land differently on its successor: a different default length, or a different instinct about when to ask you a clarifying question first. Re-run two or three of your most-used prompts and adjust once, deliberately, rather than discovering the drift halfway through something that matters.
Check the other places you pick that model. A retired entry drops out of the regenerate and branch menus on the same date it leaves the main picker, on web and on mobile alike. If your habit is to re-run an answer on one particular model, that habit goes with the entry, so choose the successor deliberately once rather than hunting for a name that has stopped appearing.
A retirement is a good prompt for an audit. If a model you leaned on for one specific job is going away, that is the moment to open the replacement's details and confirm it still does that job, rather than the moment it fails in front of you.
Why we retire models this way
There is a shortcut available to any app in this position. Leave the model name on the screen alone, point it quietly at newer weights, and let the answers change without anybody being told. Almost nobody would notice on the day, and everybody would notice eventually.
We record the retirement on the model's own entry instead: a flag saying it is deprecated, and a field naming the model that takes over. The consequence is small and it is the entire point. The model you picked keeps meaning the model you picked for as long as it exists, the successor is named rather than assumed, and when the name on your model button changes there is a reason you can go and read.
If you want to know whether something you rely on is on its way out, you do not have to follow anyone's release notes. Open the picker and read what the amber pill says.