max_tokens=700 on a reasoning model returned empty replies — hidden thinking tokens was the whole budget