During the OpenAI GPT-5.1 AMA, I noticed a lot of people complaining about one issue: the model being auto-routed. Basically, they have a preference for a specific model, but due to triggering safety mechanisms or some unknown reasons during the conversation, they get auto-routed to another model. In the Chinese community, this is often referred to as 'IQ downgrade.' Some users are so used to a particular model that they can immediately tell when they've been routed to another one. For example, there’s a user who really loves 4o, and only 4o can meet their needs. But they’ve also experienced being routed to other models and felt very disappointed. So, some people even suggested that if these models are retired in the future, maybe the parameters could be open-sourced to benefit the community. (Obviously unlikely, since it’s ClosedAI.) Regarding this issue, OpenAI staff gave a very official response, saying they’re working on developing more precise safety mechanisms that are less likely to be triggered accidentally. (Basically, they didn’t say much.) In addition, quite a few developers have reported that 'custom instructions' don’t seem to work. After setting them, the model doesn’t follow the instructions. But the OpenAI team claims that GPT-5.1 is stronger in instruction-following. Other than that, I didn’t see much broader discussion. Source: https://www.(reddit.com)/r/OpenAI/comments/1ovkt6n/were_rolling_out_gpt51_and_new_customization/