Maybe the benchmark results are different, but it certainly seems like OpenAI is doing the same with it's "thinking" step