[lxgr]

GPT-5 only, i.e. access only to the "non-thinking" model called "gpt-5-chat-latest" in the API, seems close to useless based on my experiences. Not sure if that's because of the lack of reasoning or simply because it seems to almost never pull in search results.

I've given "auto" routing a try (where some routing model tries to decide whether to route to non-thinking gpt-5-chat or the thinking gpt-5 a.k.a "GPT-5 Thinking"), and I've been very disappointed by almost all gpt-5-chat answers. gpt-5, on the other hand, has been pretty solid.

As a side note, even as an occasional buyer of CPUs and GPUs and Nokia phones in the past, OpenAI's naming scheme has got to be one of the most hostile ones in history.