Perceived Drop in GPT-5 Quality Over the Last Few Weeks

I’ll try to work around this by automating the process with shorter instructions, splitting the task into multiple steps, and validating the output at each stage. After that, I’ll leave more detailed feedback here. Still, a 30-page task should not be a serious problem for large models. Around 15,000 output tokens should not make a model like GPT-5.5 start drifting, repeating content, or losing coherence. That would be understandable for a smaller local LLM, but not for a model at this level.