Reasoning models can coaxed to reason like they do in dedicated reasoning blocks, outside of those blocks: in normal parts of the response.

But Anthropic at least has openly admitted they try to detect that and interfere