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

The model itself likely hasn’t changed. What changes are the instructions, skills, memories, context bloat, and overall harness around it. Most likely you’re hitting some form of context poisoning which can happen with too many tools, skills, plugins, etc. or if you have custom instructions. Potentially, if you have memories enabled, previously generated memories could be contributing to the problem. Usually I immediately encounter this kind of problem on any model the moment I install too many tools or skills or start adding custom instructions into the mix. I’ve found that even just having small instructions that are poorly worded or formatted (not just grammatically) can have a negative impact later on things totally unrelated to whatever the instruction was covering. I have noticed a similar drop in quality in codex as well as a weird lag every time it sends a message, which makes me suspect either something with the codex harness changed recently, or it’s just a result of bloat which is entirely possible as I do use it heavily with many skills, all of which get tossed into a list with descriptions and shoved into context I believe (so it knows it has them and when to use them.) In my case, I’m assuming it’s all the cruft I’ve built up after using it basically non-stop on weekdays. I’d say as a test try disabling some stuff and see if it improves. At the end of the day these things are just statistical next token guessers and anything wrong with the context window can eventually lead to cascading effects that just snowball especially the longer a convo goes or the bigger the file(s) it has to read.