By skyforbes I’ve recently noticed a strange behavior specific to GPT-4o when switching between models. Other GPT models (4.1, 5, 5.1) don’t behave this way.
What’s happening?
Let’s say you’re chatting with GPT-4o and you pause that conversation to test something with GPT-5.1. After a few interactions, you return to GPT-4o in the same thread.
Instead of continuing the conversation normally from where it left off, GPT-4o now does something weird:
It re-answers the last prompt you gave it before switching models, but this time with all the new context (from GPT-5.1) jammed into its window.
It’s like GPT-4o thinks:
“Oh, that last message wasn’t handled yet? Let me respond to it now… but also include everything that happened after it.”
Which ends up producing a totally different kind of reply, often misaligned with the original intent.
Before (expected behavior):
- Prompt A → answered by 4o
-
Switch to GPT-5.1 → B, C, D
-
Switch back to 4o → continue from D as normal, using A+B+C+D as context
Now (unexpected behavior):
- Prompt A → answered by 4o
-
Switch to GPT-5.1 → B, C, D
-
Switch back to 4o → re-answers Prompt A, but with B+C+D added to the context (???)
Feels like context stacking is broken or the pointer to the latest reply got misaligned, causing 4o to treat an already-answered message as “still pending,” but now with additional context that warps the intent.
This does NOT happen on GPT-4.1 or GPT-5.1—only GPT-4o shows this odd regression.
also, this post was refinded by AI, i'm sorry because my English is not advanced enough to explain this problem into an easy to understand way….