Reasoning proofs, using formal proof languages, and tool assisted math were always in their realm. The pure LLM is still really bad at the “calculation” part though. In fact it’s worse than predicted: many thought byte-level models would improve that, and… those never came to be.