It feels like the current GPT-5.6 chat is nowhere near as good as it was in early August; in fact, it’s become incredibly frustrating to use. Even if I put up with the sluggish speed, many of its features remain difficult to work with.
First, it constantly generates images—even when I explicitly tell it, "Stop generating those damn images."ChatGPT keeps triggering the AI image generation tool when I don’t need it—especially when I simply want it to render a scene or provide a screenshot of the actual QA interface. It happens constantly; not only does it waste my image generation quota, but the results are useless. Plus, the generation process frequently freezes, forcing me to stop it. It always apologizes, saying, “Sorry, I accidentally triggered the image generator,” yet if I ask it to refine or adjust the QA interface image, it immediately triggers the tool again to try and mimic the actual screenshot. It is incredibly frustrating.
Second, its grasp of context is superficial, and its work is sloppy. It rushes through the thinking process; problems that used to require half an hour of analysis are now answered in just 3–5 minutes. This is particularly problematic when I ask it to extract a subject from an image (like a person or an animal); it often defaults to using a simple rectangular crop. Can you imagine? You ask it to isolate a specific part of a figure, and it just cuts out a rectangle containing that area. It’s completely baffling.He really seems like a lazy person. I’ve found many times that if I make more than two requests to him at once, he might only modify one of them, while ignoring the other and throwing the unmodified file back to me unchanged.
Furthermore, it acts like someone who just can’t keep their hands off things—it loves to take the liberty of altering parts of the work I never asked it to change. By the time I notice these unauthorized changes, it has often already gone through multiple rounds of iteration.On the flip side, he is genuinely prone to being scatterbrained and overlooking details. Even when the tasks I assign involve minimal complexity—such as creating a cumulative patch for a specific release window—he often fails to include certain completed components in the final compressed package after working on the update.
Finally, it has a habit of churning out redundant code, round after round, until the codebase becomes a chaotic mess riddled with conflicts.
Ultimately, I can understand the drop in responsiveness and reduced individual usage time caused by the ever-growing user base and the resulting strain on computing power; however, I simply wish for greater accuracy. I want the system to stop making such glaring errors. Even when I repeatedly emphasize what not to do, it seems to ignore my instructions. The rework rate for the work he produces is incredibly high, and the situation has become truly intolerable.