OpenAI, your paying users shouldn't have to become unpaid QA engineers just to get reliable responses.
retained text
OpenAI, what exactly is happening to ChatGPT's reliability? I'm a long-time paying power user, and I'm genuinely tired of spending more time investigating bugs than actually using the product. Over the past several days, I've documented disappearing responses, premature task completion, inconsistent instruction-following, and significant differences between Instant and Medium reasoning configurations. And no, this isn't just another "the model feels worse" complaint. **I actually tested it.** # Two models. Two A/B/A tests. The same suspicious pattern. **GPT-5.6:** * Medium → Successful. Response completed and remained visible. * Instant → Failed. Generated response disappeared. * Medium → Successful again. **GPT-6:** * Medium → Successful. All requested story events completed. * Instant → Failed. Response ended prematurely. * Medium → Successful again. That's four successful Medium tests and two problematic Instant tests across two model generations. The failures aren't identical, and I'm not claiming they necessarily share a root cause. But this pattern deserves an actual technical explanation. # And the disappearing-response issue goes deeper. During my earlier investigation, Chrome Developer Tools showed: * HTTP `200 OK` * `POST /backend-api/f/conversation` * `text/event-stream` * SSE completion indicator: `complete` * The response visibly generating * The entire assistant message disappearing immediately afterward I even reproduced the disappearing-response issue using GPT-5.5 in a completely new conversation. I tested different devices, cleared cookies, collected screenshots, inspected network requests, and documented the behavior. **OpenAI Support Case: #16904787.** I've already provided extensive technical evidence to support. What I still don't have is a confirmed root cause, a meaningful engineering update, or an ETA. # Here's the question I want OpenAI to answer. **Has the effective behavior of Instant changed?** OpenAI describes Instant as the faster option for everyday tasks, while Medium provides additional reasoning effort. That's perfectly reasonable. But does this mean long-form creative writing and complex multi-step instructions are now effectively expected to run on Medium? Because that's what my recent tests seem to suggest. Medium completes my stories more reliably, but often writes significantly more than I requested. Instant tends to match my preferred length, but has shown serious reliability problems. I've adjusted my custom instructions repeatedly, and the differences remain. # This is about more than creative writing. People use ChatGPT for software development, research, professional work, creative projects, and long-term workflows. When an assistant visibly generates a complete response and then loses it, that's not a minor inconvenience. When it repeatedly stops before fulfilling explicit instructions, the user has to spend additional time checking and correcting the output. And when support communication doesn't provide meaningful technical clarity, confidence in the product suffers. I understand that debugging complex systems takes time. I don't expect every issue to be fixed overnight. **But I do expect transparency about known problems, meaningful follow-up on documented reports, and reliable basic functionality.** # What makes this especially frustrating? I've already started testing Claude, Gemini, DeepSeek, and other alternatives. Not because I wanted to leave ChatGPT. But because I need tools that reliably support my work. And yes, some competitors are producing surprisingly strong results. I'm not claiming OpenAI deliberately degraded anything. I don't have evidence for that. But the lack of clarity about these changes is frustrating. # So, OpenAI, here are my questions: 1. Have Instant's effective reasoning, stopping, or output-length configurations changed recently? 2. Are there known differences in response-finalization or persistence behavior between Instant and Medium? 3. Is long-form creative writing still an intended use case for Instant? 4. Are the disappearing-response reports being investigated? 5. Can users with documented support cases receive meaningful engineering updates? I've already submitted the evidence. I've already performed cross-model testing. I've already documented the failures. **What I shouldn't have to do is reverse-engineer the behavior of a paid AI service just to understand why it no longer reliably completes the tasks it used to handle.** I originally subscribed to ChatGPT to work on my projects. Instead, I've spent days acting as an unpaid QA engineer. **OpenAI, if your users are doing this much debugging for free, the least you can do is tell them whether you've identified the problem.** Because at this point, the competition isn't even the biggest threat to your customer retention. **The reliability of your own product is.**
document fetched 2026-10-13 00:29 · canonical https://reddit.com/r/OpenAI/comments/1x49y77/openai_your_paying_users_shouldnt_have_to_become