Eh, you're seeing raw thinking tokens. With Claude <x> 4, and I think GPT-5 series, you are no longer seeing real thinking tokens, but "summarized" tokens that are probably highly different to the raw thinking.