We’ve noticed what appears to be a regression in how GPT Apps behave on desktop.
Previously, once an app had been invoked in a conversation (for example using @appname), it remained active for the rest of that conversation. There was no need to invoke it again for every subsequent message.
Now, on both:
ChatGPT in the browser (desktop)
ChatGPT Desktop app
the app seems to lose context after every prompt. Unless the app is explicitly invoked again on each message, ChatGPT falls back to either:
a regular web search, or
its base model knowledge/training.
This significantly degrades the user experience, especially for apps designed to support a continuous conversation.
Interestingly, this behavior does not seem to occur on the mobile app, where the app appears to remain active across the conversation as before.
Expected behavior
Invoke the app once in a conversation.
All subsequent messages continue using that app until the user explicitly switches away.
Current behavior (Desktop)
The app must be invoked before every single prompt.
Otherwise ChatGPT ignores the app and responds using web search or its default knowledge.
Is anyone else seeing the same behavior? Is this an intentional change or a regression?
Thanks for the pointer! I don’t think it’s the same issue.
In our case, the app is invoked correctly on the first message and tool calls work as expected. The issue is that the app is no longer kept active for subsequent messages on Desktop, so we have to invoke it again before every prompt.
I believe the issue you’re referring to is more likely the one discussed here:
As far as I know, that regression has already been fixed. Our issue seems to be a different regression related to app persistence within a conversation.
I invoke my app (@pagesjaunes) and ask for a restaurant.
The app is correctly called.
It returns a carousel of restaurants.
Step 2
I ask: “What’s the phone number?”
I do not invoke the app again.
Instead of calling the app to retrieve the restaurant details, ChatGPT answers using its own knowledge (or web search). No tool lookup or tool invocation occurs.
Step 3
I ask the exact same question again, but this time I invoke the app.
The app is called correctly.
It opens the detailed business card for the restaurant and returns the correct phone number.
This shows that the app itself is working correctly. The regression is that it is no longer considered active after the first prompt.
I repeat the exact same workflow using the ChatGPT mobile app.
Step 1
I invoke the app.
The restaurant carousel is displayed.
Step 2
I ask: “What’s the phone number?”
I do not invoke the app again.
ChatGPT correctly calls the app, opens the restaurant’s detailed business card, and returns the phone number.
The expected behavior is still present on the mobile app, but it appears to have regressed on both the Desktop web client and the ChatGPT Desktop application.
Previously, Desktop behaved the same way as Mobile: once an app had been invoked in a conversation, ChatGPT would continue using it automatically whenever subsequent prompts required it. Now, the user must explicitly invoke the app before every prompt, otherwise ChatGPT falls back to its own knowledge or web search instead of calling the app’s tools.
I hope these conversation IDs and screenshots help identify the regression.
After running some new tests today, it looks like the issue has been resolved.
The app is now correctly kept active throughout the conversation on Desktop, and subsequent prompts continue to use the app without requiring it to be invoked again each time.
Thanks to the OpenAI team if this was fixed behind the scenes! It seems the regression has been addressed.
There might have been a regression, but apps only stay active for two turns on chatGPT webapp at the moment. After the 2nd turn the model is unable to use the tools and readding the app with @ results in an error.
We thought this had been resolved, but the app-persistence regression seems to have returned on the desktop web client. Same use case, same questions, app invoked in the first prompt only, not re-invoked afterwards.
One difference compared to when I reported this as fixed at the start of July: back then, the app was still being re-referenced automatically in the subsequent user prompts, which is what kept it active throughout the conversation. That is no longer the case at all today.
Browser (desktop web)
Conv ID : 6a6b0054-75a0-83eb-8f7c-3ecbcfe5cd85
The app is invoked on the first prompt.
It is not re-referenced automatically on the following user prompt, and it is not used in the response.
ChatGPT ends up answering via web search.
Desktop app
Conv ID : 6a6b01f7-83f0-83eb-948a-708cd3d3c5ec
The app is invoked on the first prompt.
It is not re-referenced automatically on the following user prompt, but it is still used in the response.
The app is called appropriately.
Mobile app
Conv ID : 6a6b0283-e97c-83eb-a5f4-0dcb3222890a
Same behavior as the desktop app: the app is not re-referenced automatically on the following prompt, but it is still used appropriately in the response.
Note: on mobile I’m currently seeing a separate UI-rendering issue when one of our tools is called, which can make the conversation harder to read. This is a different problem and not the subject of this thread — I’ll open a separate report on it once I’ve investigated further.
One thing worth flagging: the behavior seems quite erratic, and it doesn’t match what CharlesS described just above. On the webapp he sees the app staying active for two turns, followed by an error when re-adding it with @. On my side, the browser drops the app from the second prompt onward, and I don’t get that error on re-invocation. So even on the same client, the observed behavior differs from one setup to another.
On a more general note, and asked in good faith: could the OpenAI team share a bit about the long-term vision for third-party apps in ChatGPT?
We’re genuinely invested in building on this platform, but the experience has been unstable for several months now.
The difficulty is that these apps are surfaced under our own brand. End users don’t distinguish between the app and ChatGPT itself — so when it doesn’t behave as expected, it’s our company’s reputation that takes the hit, even when the root cause is a platform-side regression.
Understanding how committed OpenAI is to this ecosystem, and what level of stability partners can plan around, would really help us decide how much to keep investing here.
The behaviour is indeed a bit erratic as sometimes, chatGPT does a connector search and figures that he need to use the cottages MCP app, but it the app is not included in the context after the 2nd turn.
We’re seeing a very similar issue on the free tier. After one or two turns, ChatGPT appears to silently fall back to web search or an internal model response, producing results that feel inconsistent or off.
The main problem is that this behavior is not predictable or transparent, which creates a confusing experience for end users. We are still running in developer mode because we have not been able to reassure our customers that the issue is unrelated to the plugin/MCP. Even if we could, this still is not the experience we are looking for multi-turn MCP interactions, which we consider the right approach for a conversational flow.
We initially assumed the behavior might be related to the number of tokens consumed by the tool, but it would be helpful if OpenAI can clarify that or provide a way to mitigate or control the experience.
Following up again: the persistence regression on the desktop web client is still present, and I haven’t seen any response from the OpenAI team since my last posts.
Here is a fresh conversation ID reproducing the issue, where the final prompt falls back to web search instead of using the app:
The behavior is unchanged from what I described above: the app is invoked on the first prompt, is not carried over to the following user prompts on desktop web, and ChatGPT ends up answering via web search.
I’d really appreciate any sign that this has been acknowledged — even just confirmation that it’s on the radar. And the broader question from my previous message still stands: what level of stability can partners building apps on ChatGPT reasonably plan around?
