Problem Solutions
Why ChatGPT Keeps Switching Models on Its Own — and How to Lock In the One You Picked
You picked a model, started working, and somewhere mid-conversation the answers changed — shorter, flatter, subtly wrong in style. You did not touch anything. On OpenAI's own community forum, a Pro subscriber described the sharpest version of the problem: chats started on one flagship model were automatically converted to another, more expensive one overnight, with nothing in the interface making the switch obvious unless you opened the model dropdown to check. The thread's request — a model lock — does not exist yet. But most switching has identifiable causes, and each cause has a countermeasure.
The three reasons ChatGPT switches models
1. Usage-limit fallback. Every plan has limits per model. When you exhaust the allowance for the model you selected, ChatGPT does not stop the conversation — it quietly continues it on a smaller fallback model. The chat reads as continuous, but the engine underneath has changed. This is the most common cause, and the easiest to misread as the model "getting dumber."
2. Model retirements. OpenAI retires models from ChatGPT on a rolling schedule. When a model is removed, conversations and defaults tied to it move to a successor. The retirements this season have been frequent enough that developers pinning model names in tools like Codex have been hit repeatedly — each retirement date silently breaks whatever pointed at the old name. For chat users the effect is gentler but similar: the model you chose simply stops being the model you get.
3. Automatic conversion of old chats. The forum complaint that drew the most support concerns legacy conversations: reopen a chat started on an older flagship and it may be converted to the current flagship without asking. The poster's objection was practical, not sentimental — the new model drew heavily on their Pro allocation for casual ongoing tasks the old model handled fine, and the mid-thread change altered the tone and formatting of work in progress.
How to tell which model you are actually on
Open the model picker at the top of the chat. It shows the model the conversation is currently using, which is not necessarily the one you selected when you started. Check it before important work and again after any long pause or heavy session — a limit-triggered fallback usually happens at exactly those moments. If the picker shows "Auto," ChatGPT is choosing for you on every message, and variation in quality is the expected behavior, not a bug.
What you can do about it
Pick the model explicitly, per chat. On paid plans, selecting a named model from the picker — rather than leaving Auto in charge — removes per-message routing from the equation. The selection holds for that conversation until a limit or a retirement intervenes.
Turn on additional and legacy models. In Settings, paid accounts can enable options such as "Show additional models" and legacy model access, which widen the picker beyond the default Instant and Thinking choices and restore older models that are still offered. If a model you rely on seems to have vanished, this setting is the first place to look.
Start critical work in a fresh chat, model first. Choose the model before the first message rather than continuing an old thread. Fresh chats are immune to the legacy-conversion problem, and you begin with a full model allowance instead of whatever an old conversation has left.
Treat a sudden quality drop as a limit signal. If answers degrade mid-session, assume fallback before assuming the model changed. Waiting for the allowance to reset — or deliberately continuing on the smaller model for lighter tasks — beats re-prompting a fallback model as if it were the flagship.
Use the API when the model must not change. The only route where the model is exactly the one you named, every time, is the API: you specify the model per request and pay per token. It is a cost decision, but for work where consistency matters more than convenience, it is the genuine lock that the chat interface does not offer.
The honest limit
There is currently no setting in ChatGPT that guarantees a conversation will never change models. The feature the forum is asking for — a persistent lock, an always-visible model indicator, and a global default for new chats — remains a request, not a product. Until it ships, the workable discipline is: pick explicitly, check the picker, start fresh for important work, and read quality drops as limit events. That will not stop every switch, but it will stop every switch from being a surprise.