It is the reasoning. During the reasoning process, the top few tokens have very similar or even same logprobs. With gpt-oss-120b, you should be able to get deterministic output by turning off reasoning, e.g. by appending:
{"role": "assistant", "content": "<think></think>"}