“Greg Brockman commented that on one benchmark, the improvement from the previous OpenAI model to the current one was only from 98% to 99%, indicating that some benchmarks are saturated.”