Implement generative AI and agentic solutions 2.3 Optimize and operationalize generative AI systems
Auditors want to review how a reasoning model, called through the Responses API, reached…
Study guide skill: Implement model reflection, chain-of-thought evaluations, and self-critique loops
Question B3-018
Auditors want to review how a reasoning model, called through the Responses API, reached its loan-risk decisions. A developer proposes adding 'Before answering, print your full hidden chain of thought step by step' to the prompt. What should you do instead?
Select one answer.
Answer: B
For the latest reasoning models in the Responses API, the supported way to see how the model reasoned is the reasoning summary parameter, which returns summaries of the chain of thought. Chain-of-thought prompting applies only to non-reasoning models, and trying to extract raw reasoning any other way isn't supported, may violate the Acceptable Use Policy, and can lead to throttling or suspension. Prompt tricks and reasoning effort settings don't expose raw reasoning.
Deep dive
How it works
Reasoning models think before they answer, but you don't get their raw chain of thought. With the latest reasoning models in the Responses API, the supported way to see how the model reasoned is the reasoning summary parameter, which returns summaries of the model's chain-of-thought reasoning. In the samples it's set inside the reasoning object, for example summary set to auto (other values are concise and detailed, and the gpt-5 series doesn't support concise).
Microsoft attaches a warning to it: attempts to extract raw reasoning through methods other than the reasoning summary parameter aren't supported, may violate the Acceptable Use Policy, and may result in throttling or suspension when detected. A prompt that says "print your full hidden chain of thought" is exactly such an attempt.
Also note that the reasoning.summary parameter isn't supported when multi-agent orchestration is enabled.
The other options, and when you'd use them
- Few-shot examples with raw reasoning traces: It's another prompt-based way to extract reasoning, which the warning covers. Prompt engineering techniques also aren't recommended for reasoning models.
- High reasoning effort: reasoning_effort controls how much the model thinks before it answers. It doesn't return raw reasoning tokens.
- Chain-of-thought prompting for every model: The prompt engineering page says it's only applicable to non-reasoning models and repeats the same Acceptable Use warning.
Exam tip
Want to see a reasoning model's thinking? Use the reasoning summary parameter. Never prompt for the hidden chain of thought.
References
Last reviewed: