#Chats in project lose all context when switching models

1 messages · Page 1 of 1 (latest)

hollow wolfBOT
#

Reported by @zinc elk

Bug Report: Chats in project lose all context when switching models
`Steps to Reproduce`

Start a chat in a project (tested with files, and with included info) -> start on GPT-5 thinking and place at minimum 1 message -> switch to GPT-5 and place one message -> switch back to GPT-5-thinking and place a message -> await response.

`Expected Result`

Switching between models should not impact the context the model sees. it loses ALL chat context, thinking start without project context as well, only the written message shows any context of the project, but not the text.

`Actual Result`

When switching to a thinking model (o3, o4-mini, o4-mini-high, GPT-5-thinking) when coming from a non-thinking model, full chat context may be lost.

`Environment`

Windows 10, browser (chrome); Android, app; Android, browser (chrome)

#
Additional Information

Please provide relevant details to help resolve the issue, such as:

  • ChatGPT Shared Link (if applicable).
  • Screenshots or videos demonstrating the problem.

-# ➜ Need to contact support? Visit the OpenAI Help Center.

zinc elk
#

https://chatgpt.com/share/68990177-2614-8003-9ec8-94f3a629dbbd

What the clip shows:

  1. Showing chat-context about a bug I was facing when running HPO trials in my code, my message -> GPT-5 answer -> (1) my message -> (2) GPT-5-thinking answer
  2. for (2): Showing an earlier answer, where the thinking did not take in chat AND project context, while the message itself did take into account project context. can be seen that the thought process has information about crypto currency while I've legit never mentioned it (I've seen this default behaviour about tether multiple times now, if it doesn't have any context)
  3. Resending my (1) message (instead of retry because it ISNT AVAILABLE IN PROJECTS, second bug...) on GPT-5-thinking, getting (2) back with thought that show it is ONLY getting project context, but ALL chat context is lost. Eventual message also shows complete loss of chat context.
ChatGPT
ChatGPT - Memory error in optuna

Shared via ChatGPT

#

The following clip shows that if you use GPT-5;
Reseding to get a new answer gets you the following:
GPT-5, automatic "Thinking longer"

  1. The thinking clearly shows a loss of context, and the eventual answer back also shows that all chat context is lost, project context is still there.

GPT-5, automatic "Thinking longer" -> "Quick message"

  1. Gets an immediate answer WITH Chat AND project context
  2. No loss of context, works as intended and how GPT-5-thinking should also work!
#

I want to make it especially clear; this bug ALREADY was there when o3 and o4-mini-high were the models of choice, I believe this has nothing to do with the model picker, but something inherently wrong with the Project feature (just to reduce the scope a bit!)

storm leaf
#

I rely on this feature too. Didn't we discuss this in another chat or report somewhere?
But you've definitely taken this report to the next level.
Did you open a report with OpenAI about this?

zinc elk
#

So many people are now getting this issue because GPT-5 automatically can switch to thinking, and switching to thinking specifically in Projects is broken

#

Since before the GPT-5 update, the demographic of people who used Projects+ used thinking models + actively switched in long conversations to thinking models was just me.

With the GPT-5 update, the thinking usage has more than tripled, and the automatic switching in GPT-5 makes it way easier to encounter this bug.

zinc elk
#

Reddit has been full of posts like this, but under the hood it's an issue with Projects, not with GPT-5

#

It's at least for public perception a huge PR hit just because of a bug in a small feature

storm leaf
#

OpenAI seems to be pulling the trigger on every footgun they can find. SMH They'll recover eventually.
Reddit and Discord are places to vent, not to get remediation. If you've reported the issue to OpenAI via Settings > Help > Help Center then you've done your part. But now we collectively need to nurture that process...

#

Support people at any company are just people with varying degrees of knowledge, insight, caring, dedication, etc. THEY are asking YOU for network data?? That's nuts. They have all of the network data. They just just ask you for a chat URI and they'll have all the data they need to trace exactly what happened in that series of transactions.

#

Be cordial, positive, collaborative, but firm, and work with them to understand the problem and the severity of the issue.

#

They probably aren't getting too many Good reports on this because it's tough to provide qualitative evidence of the issue - I'd think most people would just go "hmm, that was weird" and move on, or complain and move on...

#

I haven't tested this one myself but I'll try as time permits. Most of my time as a developer is spent just trying to get things done so I can't commit to experiments. But I promise you that if I see this in action now, I'll report it too and walk it through with Support. That's all we can do.

#

Thoughts?

zinc elk
#

I've worked both in technical support for companies and as a technical tester for website and app, and I can confidently say that a good pipeline makes all the difference here really, I think they just haven't figured it fully out yet.

In this case, a clear bug report like this should obviously be recognized by a support agent, but they're probably using some cheapo support from what I've seen. Every support message feels written with 4o and apersonal.

It kinda makes sense, they explicitly asked for a de-anonomized HAR file, I gave them that+a complete analysis of what I thought was going on here, just for the support agent I guess. Maybe it's easier for them to see signature etc from my local network calls instead of on their end, they might not save all packets tbh.

#

I've tried report this three times now, and the first 2 times I finally got a support agent, they just give me random "workarounds" or "tips on prompt engineering" which are utterly useless

storm leaf
#

Yeah, we've both been in the same chair. You know the personalities, limits, etc. Work within the limits and then attempt to push past them. The chatbot needs a clear and concise report that's worthy of human attention. When you get an accurate confirmation from the bot, be sure to tell it that you now want the report sent to a human Support rep. (Oh what a new world that we live in...) Support should respond within days.

#

Tier 1 Support at any company will ask the same silly questions even if they are answered in the OP. That can be from a lack of concern, lack of knowledge, or just a desire to move on to the next ticket and claim a job well done. Be sure to include your diagnostic results in the initial report to try to preclude that, or just for the right to say "already answered that in the first exchange".

#

Once you get through to a human rep I think the bot is bypassed in the pipeline. If you're still chatting with a bot, you haven't made the case.