OpenAI uses substantially different model for subs vs. API?

Yeah, my colleague recently checked it. He saw that thinking tokens for same model and same thinking level are about half for subs vs API.

What?