I think max_tokens restricts both the reasoning and response tokens. That means if max_tokens is reached in the reasoning, the rest of the reasoning and the response will be cut off. Maybe thats what is happening in your case.