GPT5’s performance is disappointing. It frequently loses context and tries to turn all responses into lists. Even for the simplest home electrical DIY questions, it repeatedly emphasizes safety warnings about wearing gloves but provides no substantial guidance. When discussing emotional topics, it only mechanically guides through breathing exercises before forgetting the original question.
As a heavy user of programming and mathematics, I have to frequently switch back to GPT4o. The difference between them is like drinking aged fine wine versus sewer water. OpenAI claimed it would remove the model selector, but now paying users have to constantly switch manually — this experience is terrible.
The root of the problem lies in costs. GPT5’s server architecture has clearly been downgraded, sacrificing long-term memory and processing capabilities. What’s more ironic is that even GPT4.1 performs better than this so-called new flagship. Some users have noticed that GPT4o’s output quality is now starting to fluctuate, suspecting that the company is secretly testing hybrid models.
The most tragic part is the entire system design: multiple AI models form a virtual customer service team, unaware of each other’s conversation history, with only one having internet access, while all models are busy avoiding legal risks. This essentially turns ChatGPT into a simulated dysfunctional corporation.
When technological progress yields to accounting books, even the best technological visions become user experience nightmares.
Link:Source
