“OpenAI's Deep Research product is an existence proof that reasoning models can succeed in domains without easily verifiable metrics for success.”