Version numbers go up. Check what changed before you switch.
A newer AI model can drop a feature the older one had. Here is what to check in MultiChats before you switch models.
A new model lands and the version number climbs. That number records one thing: a lab finished a training run and shipped it. You still need to check whether the result suits your work.
As of 19 October 2026, MultiChats runs 59 models from 17 providers, and new ones arrive most months. About a minute of checking tells you whether a newer entry fits the thing you are doing today. Here is where to look, and what tends to change between one generation of a model and the next.
Capabilities can move backwards
A new generation sometimes arrives carrying less than the one before it.
Google's Flash line shows it plainly. Gemini 3 Flash takes images, PDFs and video. It searches the web, exposes thinking control, and works in voice calls in the mobile app. Gemini 3.8 Flash, added on 4 October, matches every one of those except voice.
supportsVoiceMode reads true on Gemini 3 Flash and false on Gemini 3.8 Flash. The note recorded beside it: no verified realtime voice path on the newer model.
The same pair splits again on thinking. The effort control on Gemini 3 Flash includes an Off setting. On Gemini 3.8 Flash it does not, because Google's API has no setting that makes Off reachable on that model. By default the app picks the effort from your message; on web the control appears once you turn Intent Detection off in Settings, then Tool Preferences.
On mobile, a voice call can run on the chat model you picked. Choose 3.8 Flash there and the app offers to move you to Gemini 3 Flash instead. Version numbers alone would never tell you that.
MultiChats puts the leading AI models in one app. Try it free
The same trap inside one family
Generation gaps are the obvious case. Sibling entries catch people just as often.
xAI's Grok 4.3 appears twice in the picker, once with reasoning and once without. Grok 4.3 accepts PDF attachments. Grok 4.3 Reasoning refuses them, because supportsPdfUpload is false on that entry and true on the other.
Attach a PDF to the wrong sibling and the composer blocks it before the message goes out. The two names give you no warning at all, and they sit next to each other in the list. The capability grid on the model info card settles it.
What the New chip actually marks
The picker paints a red New chip on recent arrivals. It reads an addedAt date, and that date records the day a model landed in MultiChats.
Release dates run on a different clock. Nineteen of our current entries share the same addedAt value, 9 August 2026, because they went in as one batch. Their makers shipped them across half a year.
Our Recommended tab is a hand-set flag, model by model, so it does not track version numbers either. Gemini 3.5 Flash Lite carries it and the older Gemini 3.1 Flash Lite does not. The pair splits the other way on voice: the older one works in voice calls, the newer one does not.
A tuned prompt is tuned to one model
Anyone who uses AI daily ends up with prompts that work. The phrasing, the ordering, the amount of context you paste: all of it settles around one model's habits.
A new version changes those habits. It may write longer. It may front-load caveats you had trained out. It may restructure a format you had spent a fortnight pinning down.
Those changes read as a quality drop, which is what makes them easy to miss. The prompt still runs. The output now needs editing it did not need last month, and the editing is where your time goes. Rewrite the prompt for the new model, or stay where you are, but decide it on purpose instead of drifting.
Speed is part of the answer
A newer model can also be slower to start answering than the one it succeeds. On a phone, held one-handed on a train, that wait changes how the whole tool feels.
Thinking models are their own subject. Some tasks reward patience, and plenty of the things people ask an AI every day do not. If a model answers while you are still reading your own question back, that responsiveness is worth weighing against whatever the newer entry does better.
Launch benchmarks are a chosen snapshot
Every launch arrives with a table of scores. The company that trained the model picked the tests, ran them itself, and published on the day it most wanted a good number.
Nobody is cheating. The tests are chosen, the timing is chosen, and neither choice was made with your desk in mind.
Your own test takes five minutes:
Take a real piece of work you did last week.
Run it on the model you use now.
Run it again on the newer one.
Compare the two answers you would actually send.
That comparison beats any scoreboard, because the work is yours.
When a model you like retires
Retirement is signposted here. A retired model is marked as deprecated and given a named replacement, and the app shows you both.
What you actually see:
A Deprecated chip with a warning triangle on the model row, in the web picker and in the mobile one.
On web, hovering that chip on a model your plan includes gives the discontinuation date, when one is set, and the replacement's name.
A warning block on the model info card reading Deprecated and "May be removed soon".
On mobile, deprecated rows sorted to the bottom of the list.
GPT-5 is the live example. It points at GPT-5.4 as its replacement and carries a discontinuation date of 11 December 2026. The row stays in the picker until then, with the warning on it. Picking it does not get you GPT-5. A deprecated model is redirected to its replacement on every send, from deprecation rather than from the discontinuation date. After the date the row disappears.
Old replies keep the name of the model that wrote them, so your history stays accurate about what answered.
Three checks before you switch
Switching takes one tap and reverses with another. Pick a model while you are inside a conversation and that same thread continues with it from the next message on.
Before you commit, read three things:
The capability icons. Vision, PDF, Video, Web search, Reasoning, Voice. The card lists only what the model actually has.
The plan. Some entries are open to everyone and some sit behind a subscription, and the picker marks that on the row.
The privacy note. We do not train on your conversations. What the provider does with them varies by model, and the models where we know of a difference carry a notice on the card. Most cards carry none, so treat it as a flag when it appears.
The short version
A version number is news. It reports what a lab shipped, on the lab's schedule, for the lab's reasons.
Your question is narrower. Does this model do the thing you need, at a speed you like, on the plan you have? With nearly 60 names in the picker, the answer is often one that has been sitting there for months. Choose it for the capabilities you need.