The way almost all modern LLMs work is by starting with a "thinking" phase which is explicitly not part of the output. You let that process run longer.