OpenAI releases under the *same name* models that work completely differently — some where most of the lifting is done by an LLM, others where most of the lifting is in test-time CoT search. This can be confusing. But the multi-min latency is how you can tell them apart.