They are all using mid-tier gpt-5 variants (not the "-high" one that's hidden by default, not gpt-5-thinking) and don't realize it