Environment
- Surface: Windows 10
- Browser or app version: Firefox developer edition (portable)
- Affected feature and model, if relevant: 5.6
Greetings. I’m new here and this is my first post, but I’ve got a lot on my mind—so to speak, things have piled up… I just want to vent a bit and hear the community’s opinion.
This post about ChatGPT issues probably won’t offer anything new that hasn’t already been discussed here, but it might serve as a summary of the problems, since I’ve experienced them firsthand rather than just hearing or reading about them on a forum.
I’ve been using ChatGPT since early 2026 and have witnessed the evolution and upgrades of several model versions; generally, these changes were improvements. However, version 5.6 has been the absolute worst experience for me so far. The problems intensified about a month ago—coinciding, I believe, with the release of “Astra.” It has reached a point where I’m asking myself, “What am I paying for?” and considering trying a different model. I don’t know the exact cause of these errors—whether it’s “optimization,” issues stemming from the new model launch, or a deliberate nerf. A conspiracy theory is starting to cross my mind: perhaps the company decided that Plus subscribers were doing too much, so they needed to cut back on capabilities.
Before describing the issues, let me briefly explain my workflow. I conduct research projects structured like this: I feed the model a detailed technical brief designed so that it can essentially operate autonomously based on that input—in most cases, my only response is “continue.” I run many of these research projects simultaneously, keeping multiple browser tabs open for them.
- The main issue making work nearly impossible is the failure of the Python environment within the model. These are errors of the type
caas.internal.errors.ClientError(sometimes simplyClientError). When this error occurs, the model cannot perform any useful tasks; effectively, the chat is reduced to “casual conversation” and basic web searches. Crucially, the failure can occur at any of three levels:
- the internal Python runtime
- the container/shell + python3 environment
- the user-facing Python runtime
These three components do not necessarily fail simultaneously; they can function (or fail) independently. I am not sure exactly how they differ, but recently, my chat sessions have started with the prompt: “Check if your environment is working.” However, even if the model reports that “everything is working,” that doesn’t guarantee success. The issue is highly unpredictable, manifesting in various ways: - It works in one project folder but fails in another.
- It works fine in several chats within a single project but fails in a new chat created in that same project.
- It initially reports a failure but works correctly after the response is regenerated.
- Within a single chat, a series of responses goes smoothly, but after a pause, it stops working.
Additionally, when creating multiple chats, a “TooManySessionsError” sometimes appears, suggesting an artificial limit on the number of active Python sessions. I had never encountered this error before.
-
“You’ve reached the maximum length for this conversation, but you can keep talking by starting a new chat.” Technically, this isn’t an error but simply a notification that the current chat has hit its limit. However, the capacity of these chats seems to have shrunk significantly of late. Roughly speaking, while a saved dialogue used to take up 800–900 kilobytes of text, a similar dialogue now occupies only 80–100 kilobytes. If we measure in terms of responses, the dialogue used to hit a limit at around 60+ replies, whereas now it cuts off around the 10th reply. Obviously, these figures are relative, but the fact that a limit exists is clearly noticeable. The only positive change recently is that, following the interface update, such chats are now marked with a red icon in the browser.
-
The next issue has been discussed at length in a neighboring thread, so I won’t go into too much detail here. It involves the model’s tools failing after an interval of 23–26 minutes. Previously, I had seen instances where the model would “think” about a response for 100 minutes or more. For my purposes, I even had to build a script that the model triggers during a response to track the elapsed time; once the ~20-minute threshold is reached, it stops “thinking” and outputs the answer immediately to prevent it from being lost due to tool failure. If memory serves, the model version current last winter behaved the same way—its tools would cut out at the 23–26 minute mark.
-
This error has generally always been present, but its frequency has increased recently: the response simply “dies.” That is, there is no specific error message in the chat—like “stream error” or “ChatGPT stream recovery polling timed out”—the model simply stops “thinking,” and the prompt input box becomes active again. Depending on the situation, the chat window might show the reasoning process it had started, or it might show nothing at all other than the user’s last prompt.
-
As I understand it, they updated the browser interface this past weekend (September 26–27) making it more app-like. And now, it seems the “Unable to load GPT conversation” issue has migrated here from there — meaning you try to open a chat, but it simply won’t load. Previously, this problem was limited to the app, but now it’s happening here too. By all appearances, it’s a “too many requests” type of error — you get throttled when trying to open multiple chats at once. It used to be standard practice for me to open or restore a browser window with a bunch of tabs open, but it seems that functionality has been curtailed.
-
This error seems specific to me. I use Firefox, and browser notifications alerting me that the chat has generated a response have never worked. No matter how many times I’ve tried changing settings, clearing cookies, and so on, nothing has changed. Even with a fresh login, tweaking the notification settings yields no results — the browser never prompts me to enable notifications for the site. The only time it ever worked was when the model executed a scheduled task; after the response appeared in the chat, a clickable control showed up. That triggered a browser prompt for notification permissions, and they started working after that. However, that functionality vanished after the recent UI update. Notifications work fine in the app, so the issue is specific to the browser.
I also want to specifically mention those so-called “checks”—messages like “Our systems are thinking a bit more about this request before responding. You can retry with a faster model for a quicker response, though it may be less capable of handling complex requests.” Not only does this significantly increase response times (from a couple of minutes in version 5.5 to over 10 minutes in 5.6), but the system also seems to “flag” you as a violator, applying these checks to absolutely all your prompts — whether in the current conversation or a new one. And that’s not to mention instances where no response is generated at all due to “cyber-censorship,” even though the content contains nothing prohibited. The situation was even worse months ago: when a response was “censored,” there was apparently a bug (or a feature?) where the model would lose a chunk of context — upon regenerating the response or continuing the chat, it would no longer “see” its previous reply, the attached files, and so on. It seems they’ve fixed that now, though (has anyone else noticed this?).
So, that’s about it. Don’t get me wrong—overall, I’m actually happy with the capabilities ChatGPT offers—but lately, it’s been a bit of a letdown (especially when you compare it to version 5.5). What’s really annoying is that these aren’t just the usual “yeah, we know something’s broken, please bear with us” issues; it’s completely unclear what’s actually going on—like, is this intentional? When will it be fixed? And so on.
Built-in Python runtime — failed immediately with