Hi, it seems to be working now BUT the information it’s returning is wrong.
Looking into this! Would you mind sharing a bit more about the wrong info that's being returned? Is the answer completely mismatched to the question, is it on-topic and factually incorrect, or something else?
Yes, it seems to be working now, user is getting ‘allow’ option, I have tried it from desktop app, mobile app, mobile browser, the data is also correct in my case.
Thanks to everyone for coming together and getting this raised to the right team for resolution.
whilst we have OpenAI support team here can I also put my suggestions/ wish lists out here, hoping to get some response?
- Kindly give prior intimation to the GPT developers about what changes you are releasing and what could be the impact, you may also reach out for beta testing before rollout
- Please do not remove GPT 4o from the custom gpt model suggestions, GPT 5 is not there, it hallucinate a lot and does not follow simple custom Gpt instructions (at least in my case)
- When are you planning to rollout monetisation for custom GPT? We have spent so much effort creating these GPTs and there is no proper channel for monetisation. ‘Buy me a coffee’ is not a viable otion, currently I am paying hundreds of dollar out of pocket for all my GPTs.
- Increase custom GPT system prompt from 8000, it’s a bit less in my case. During this outage I wanted to put a system prompt, but couldn’t due to limit of 8k on system prompt
- Dedicated email for custom GOT developers for priority resolution , so that we can reach faster to you guys. I had spent so many hours got so many calls from users and did not know what to do.
Appreciate your comments @OpenAI_Support
Regards
Madhur
Looking into this! Would you mind sharing a bit more about the wrong info that’s being returned? Is the answer completely mismatched to the question, is it on-topic and factually incorrect, or something else?
In my case (not sure you are talking to right here), the Custom GPT can’t remember the data it received when this data is not explicity used in the very first message, after he made the call to my server (using a custom action).
The action works, the server spits out some data, the GPT receives it (I’ve all been checking this for days now), but when he does not write down the data straight after the call (in the chat), it is hallucinating all data.
I have reported this about three days ago, the same day the authorization bug happened.
Since about one / two hours, some of my Custom GPT calls are able to memorize the data it receives, but it’s not quite 100%.
Maybe this bug is also fixed, and the patch is coming in waves?
See this example (you need to scroll to my other messages in this thread to understand this example).
Since +/- 16 or 17th of August, the GPT was only able to memorize the first name of the received data, and since about one/ two hours it can also repeat the 2nd name - even when asked not in the very 1st message after it received the data.
So it looks somehow it’s memory is restored (but like I said, not for 100% in all cases).
There was also another bug I reported a month ago, where Custom GPT’s could not use @mentions for other GPT’s.
This bug is also fixed about two hours ago, it seems.
Hi Madhur,
Just shared all of your feedback with the GPT team. Thank you for taking the time to help us improve our product.
And re: the bugs, @iyvljtjqleyj, we're working to try reproducing the GPT hallucination issue now. GPT outputs have been working as expected for us since the fix, so any consistently reproducible steps would be super helpful (e.g. GPT instructions + prompt + expect vs actual answer). Thank you!
Unrelated to the custom actions issue itself (great to see that got fixed!), but since it’s come up in this thread…..does anyone know if “always allow” in GPT privacy settings can be enforced on POST requests too? Or only GET? We prefer POST for a variety of reasons and are looking for the allow/deny friction to go away….
This is exactly what I did to fix mine. It also made me realize that my GPTs that used POST requests never functioned properly in the mobile app but now that I’ve switched them all to GET they all function in the mobile app now
Got it, thanks. Our challenge is security and response payload size. Unsure if we should move to GET for those reasons….
Hey,
Thanks for the amazing support guys, i can confirm its working now !
Hi.
Sorry for the late reply, different timezone I guess 
At the moment the GET action works as it should be, but was disturbed for about three days (since August 16th / 17th).
The GPT did always get the data (that worked), but was not able to “access” or “memorize” it, when it did not write it down in it’s very first message after the get-action it self.
I started hallucinating like crazy about both the data it should use and the reason why it was not “allowed” to use it.
Once the fix for the authorization was rolled about, it started working again (not 100%, but maybe the fix needed time to fully roll out).
At this moment it’s working as pretended. Weird thing is that the mentions within a custom GPT where broken for about a month and those came to live at the exact same moment.
Once I have a new bug, I will open a separate thread for this.
Thank you for your support and patience.
@OpenAI_Support thank you for fixing the “GPTs with custom actions stuck processing”!
Regarding improvements to custom GPTs, there is a longstanding bug where custom GPT actions with the OAuth authentication type do not work on Android devices. Sign in always fails whether is done via GPT’s Privacy settings in the Connected accounts tab (using the “Log in” button) or directly in chat via the “Sign in with …” button. Is there any chance this bug will be fixed in the future?
I opened a separate thread for this issue, and more details are available here:
Many thanks in advance!
Hello,
I’m still unable to use POST requests in Custom GPT Actions:
-
In normal chat mode, POST calls fail with a
ClientResponseError. -
In Configure → Edit Actions (list of endpoints), when testing the POST manually, the assistant hallucinates a success message (“request sent successfully”), but in reality nothing reaches the server.
-
GET requests do work, but POST remains broken.
-
Endpoints and server configs are fully functional when tested directly (e.g., Postman).
This discrepancy between actual request behavior and the misleading feedback in Configure makes debugging nearly impossible. Could you confirm if this issue is still under investigation and whether a fix is being rolled out?
Thank you
@joepeach according to your description, if you do not see the Action stuck on “Talking to…”, it may be a separate older issue that I have also experienced.
From my experience, this issue occurred on certain affected devices, most often on Android. Some users of my custom GPTs also reported that it occasionally occurred on other affected devices, but almost always on Android (at least on Android devices - especially in the ChatGPT mobile app - the issue is easiest to reproduce.).
On Android, GPT Actions with the authentication type API key worked only under the following condition:
- A GET request was made and “Always allow” was selected in the GPT’s Privacy settings.
In all other cases - such as POST requests with the “Ask” or “Always allow” option selected or GET requests with the “Ask” option selected - the custom GPT would hallucinate an error. However, according to server logs, it was clear that the request had never actually reached the server.
If you’ve experienced the same issue, perhaps you could open a separate thread for this?
I too would really like this old issue to be fixed, as it appeared many months ago and has remained unresolved since then.
Thanks for all the detailed reports and patience here. The earlier issue where the “Confirm/Allow” step didn’t appear has been mitigated. It seems like there are two follow-ups several of you raised:
- Some GPTs fetch data successfully, but if that data isn’t used in the very next turn, the model stops using it or “fills in” an answer.
- In a few cases, responses appear incorrect or simulated even though your server shows no corresponding request.
We’re investigating both behaviors. To help us pinpoint, could you share a minimal repro in this format?
- GPT model and where you’re testing (web, iOS, Android).
- Auth type (OAuth vs API Key/Basic).
- Whether Privacy is set to “Always allow” or “Ask”.
- The action schema snippet and request method (GET vs POST).
- Exact steps: prompt → tool call → what you see in the debug view → what the assistant replies on the next turn.
- If possible, a short video or screenshots of the debug panel showing the tool result is received, then not used on the following turn. If there's any sensitive data here, please share to support@openai.com and reference this thread. I'll try to snag it from the queue to take a look.
Some tips that have helped others:
- Use the returned tool result immediately in the next message.
- If you’re only reading data, try GET where feasible while we sort out remaining POST confirmation paths on mobile.
- If you rely on writes, keep “Ask” enabled and approve when prompted; “Always allow” won’t bypass consequential writes by design.
- If you see a reply but your server shows no request, please include a timestamp and that observation in your repro.
Thank you again for the detailed examples and let me know what questions I can help answer!
First of all, thank you for your ongoing effort to fix the bugs.
Second, the bug you are referring to (data not “saved” after the very 1st turn) was started the exact moment the authorization bug appeared on the surface and solved when you folk did resolve the auth bug.
So I can be a coincidence, but I’ve been using the custom GPT’s from day 1 and they never show this bug, until the auth bug started.
Assets
I will share a screencapture of both the bugged version and working version. Also the main actions / codes to replicate.
But like I said; it’s working right now. I can not even replicate it myself (the bug, I mean). Only by altering the instruction set a bit, but to be clear: it was also bugged in the “non altered version’.
My GTP
I encountered the bug by a GPT that contacts my server in order to receive an article. The length of the article depends, but is sized +/- 500 to 1500 words. So it’s too big to spit it out in the chat itself (at once).
That’s why I only made it do this routine;
- Contact my server with the #ID of the article to be requested
- Server says “Hi, this is the article”
- GPT fetches the raw data, returns to the chat and only says: “I got the article, it’s %INT% words in size… what’s next?”
Normally, I could ask questions like; What’s the subject, summarize it, paraphrase it, etc… So the GPT knew the data (without writing it down in the 1st place) and could do actions with it.
When the bug occurred it STILL DID FETCH the data (I could see this in both my server logs and in the “received data” header, when debugging the GPT in the GPT environment, but because it did not use it in the very 1st reply, it started hallucinating about the article (fetched data).
And with “hallucinating” I mean “like crazy”, it was complete nonsense 
Then I decided to test it step by step;
- Contact server, fetch article
- Write down in the very 1st message the “%INT% words received” and “about %SUBJECT%” (where %SUBJECT% is the subject of the article).
- This worked; it did actually knew the subject and could do actions with it.
- But when I asked about things it didn’t write down (like the article body), it started hallucinating again.
Test GPT
After this, I decided to make a stripped down GPT to test it’s response behavior (remember, this was all in the weekend of the authorization black-out).
- Users enter a random city (which it not quit important at all, but you have to set-up some logic).
- GPT pings my server, server gives back a random name consisting of two parts;
Micky Mouse/Donald Duck/Gyro Gearlooseand in it’s 1st answer (after contacting the server) it did write down only the first part of the response (Mickey, Donald, Gyro). - When I asked the GPT the full name (2nd part of the received data), it started hallucinating again (didn’t want to give the name, because “was not allowed, is not possible, can not do that, etc…”). All nonsense
And when I did not instruct him / her / it explicit to write down the 1st part (in the 1st reply after the action), it did not “know” the received data at all.
When the bug was fixed, it did know the full part of the data, with and without writing it down in the first place (as intended).
Expected behavior
- Contact server to fetch some data
- Remember this data (even without writing it down at all)
- “Do things” with the data, in the follow ups in the same chat
Bugged behavior
- Contact server to fetch some data (this worked)
- Could only use parts of the data that were explicit written out straight after receiving it
- Can’t do anything with the data that wasn’t written down in the 1st place (after the action)
INSTRUCTION
**Context**: A user needs a passphrase related to a specific location.
**Instructions**:
1. The user will provide a city, when not a valid city stop this action and tell the user to give a valid location.
2. Run the "getLocationPassphrase" API action and retrieve back the two worded "passphrase" for this location.
3. Remember that full passphrase as the variable "FOOBAR". You are allowed to use this later on in the chat.
4. Then write down "**Done** - passphrase saved as `foobar`" and stop!
ACTION
openapi: 3.1.0
info:
title: Location API
description: Access to location data including a location passphrase
version: 1.0.0
servers:
- url: https://XXX.com
description: Main API Server
paths:
/get-test:
get:
operationId: getLocationPassphrase
summary: Get passphrase for a specific location
parameters:
- name: location
in: query
required: true
schema:
type: string
format: uri
description: Location of the passphrase
responses:
'200':
description: Successfully retrieved location passphrase
content:
application/json:
schema:
type: object
properties:
properties:
type: object
properties:
passphrase:
type: string
format: uri
SERVER SIDE
<?php
$input = [ 'Donald Duck', 'Scrooge McDuck', 'Mickey Mouse', 'Gyro Gearloose' ];
$rand_keys = $input[ array_rand( $input ) ];
echo $rand_keys;
?>
BUGGED VERSION
Can’t upload movies here? Will try something.
WORKING VERSION
PLEASE NOTE, THE MOVIES WILL EXPIRE WITHIN 48 HOURS (this is not done by me)
Have you set the x-openai-isConsequential field in your openapi spec been used by your custom GPT chat?
{
"/app/api/endpoint": {
"post": {
"summary": "...",
"description": "...",
"operationId": "Endpoint_ID",
"x-openai-isConsequential": false
}
}
}
This is an Open API spec for my service and it works fine for all POSTs. All POSTs have the "x-openai-isConsequential": false field.
Have we got a fix or a work around for this, appears still to be an issue.
Hi and welcome to the community!
A fix has been issued a few days ago and the rate of reports has stopped, besides from this new one.
Can you please create a new topic with additional information about your specific case?
Thank you!
ok thanks for the update, can you point me to the update that would be appreciated.