Yes. I’m a Pro user and I’m seeing a very similar regression, but in my case it goes far beyond High simply replying too quickly. I use ChatGPT for long-running technical project work involving large archives and source trees, code analysis, debugging, multi-step changes, rebuilding and validating project files, and creating downloadable outputs. For the past few weeks the reliability has been absolutely terrible across multiple completely unrelated Projects and conversations.
The Pro execution/runtime path itself also appears fundamentally broken for me. This is not just a case of the model giving weaker answers. Complex tasks that require sustained execution, inspecting files, modifying code, validating results, rebuilding outputs and then physically checking the resulting artifacts routinely fail somewhere in the middle. A task can make substantial real progress for tens of minutes or longer and then the execution layer simply dies.
When that happens, tools may stop returning usable results entirely, execution can time out, “Thinking Failed” or “Something went wrong” can appear, tool calls can stop completing, or the run can finish without the files or outputs it was supposed to create. In some cases even very simple sanity checks stop working once the execution path has failed. The conversation may still look alive and continue producing text, but the actual Pro execution environment behind it is effectively unusable.
I repeatedly get workflows that make real progress for tens of minutes or longer and then suddenly collapse with “Something went wrong”, “Thinking Failed”, “Connection interrupted”, message delivery timeouts, or other failures. I’ve also had the execution/tool channel die in the middle of real work: container/Python/tools stop returning usable results, trivial echo/print sanity checks fail, or the UI returns things like “Skipped 1 message” instead of the expected tool result. In other cases the Code Interpreter session expires, download status fails, or the conversation claims that files/results have been completed but a later physical check shows that the expected artifacts are simply not there.
File handling has also been unreliable. I’ve had newly uploaded ZIPs and other files visibly present in the conversation UI while the execution runtime could not access them at all. I’ve seen runs reach a supposed checkpoint and then become impossible to verify because the runtime is gone, which means I have to discard work and fall back to the last state I physically verified. This makes any serious multi-step development workflow extremely risky because I can’t trust that “completed” work is actually durable.
Tool routing/orchestration is broken for me as well. ChatGPT constantly triggers image generation during completely ordinary technical and text conversations where no image was requested — sometimes dozens of times a day and on particularly bad days close to a hundred times. Even after I stop it and explicitly say “DO NOT GENERATE AN IMAGE, JUST ANSWER IN TEXT”, it can immediately trigger the image tool again. The image usually does not even finish; the run collapses and I then get a noticeably shallower, worse text response, almost as if the request has fallen onto a much weaker execution path.
I’ve also had completely benign work routed into additional safety/processing states, frontend failures such as React #185, runs that claim work is continuing when it clearly is not, and inconsistent UI/backend states after reload.
I even had another fresh failure where my message appeared normally in the conversation, ChatGPT returned “Hmm…something seems to have gone wrong”, and after I reloaded the page my own message had disappeared from the conversation history as if it had never been sent.
Another major problem is that a failure late in a long technical run can invalidate the entire session. If the runtime dies after substantial work has supposedly been completed, I can no longer trust summaries such as “the files were created”, “the changes were applied”, or “validation passed” unless I can physically inspect the resulting files myself. I have already had situations where reported progress or completed outputs could no longer be verified after the execution environment failed. For serious development work, that is disastrous because the only safe option is often to throw away the unverified work and restart from the last known-good state.
The important part is that these failures are not confined to one old conversation, one project, one browser tab, or one type of task. I have reproduced different parts of this across unrelated technical projects and fresh conversations. Sometimes one part of ChatGPT still works while the execution/runtime path is completely broken, which makes this look much more like a broader reliability/routing/runtime problem than simply “the model got worse”.
I already have an open Support investigation and have sent them extensive screenshots, timestamps, conversation links, controlled comparisons, and examples from multiple projects. Support has confirmed that the issue remains unresolved and that the evidence is part of a technical investigation, but I still have not received a concrete technical status or explanation.
At this point I simply cannot trust ChatGPT Pro for serious development work. Every time I ask it to inspect files, make code changes, rebuild an archive, validate the result and give me the finished files, I have to assume that somewhere in the process the reasoning, runtime, file access, tool result channel, response delivery, or persistence layer may break and invalidate the work.
So for me this is not merely a quality regression in GPT-5.6 Sol High. The Pro execution environment itself has become unreliable to the point of being functionally broken for serious multi-step technical work.
