“Models that perform test-time search, such as O1 Pro, have significantly higher latency, taking up to 10 minutes to answer a query, and are more costly to run.”