Bug Report: Non-reasoning models losing all context early in conversations
`Steps to Reproduce`
Just have the kind of conversation that wouldn't make GPT-5 automatically use reasoning (creative writing or casual nerd-outs worked for me). Preferably include multiple file attachments throughout the conversation (not necessary, but it makes this bug easier to notice). Once the AI suddenly switches writing style, ask it something that requires referring to previous parts of the conversation (preferably from messages close to before it changed writing style).
`Expected Result`
For the AI to remember correctly. After all, if it had to do with the context limit, it should only forget the earliest parts and not all at once.
`Actual Result`
It makes blatant mistakes. And when I mention this to GPT-5, it tries again with reasoning active and does get everything right. Showing that this only affects non-reasoning models. (The latter part obviously won't apply if you test this with GPT-4o because it can't use reasoning)
`Environment`
Tested in the ChatGPT website and Windows app.
main urchin