Implement generative AI and agentic solutions 2.3 Optimize and operationalize generative AI systems
A GPT-5 series reasoning model gives correct but overly long answers.
Study guide skill: Tune generation behavior, such as prompt engineering and adjusting model parameters
Question B8-026
A GPT-5 series reasoning model gives correct but overly long answers. You want shorter output without changing how much the model reasons. Which parameter should you adjust?
Select one answer.
Answer: A
verbosity is a GPT-5 parameter that gives more granular control over how concise the output is. reasoning_effort controls how much the model thinks before answering, which the stem wants to keep. Reasoning models other than GPT-6 Astra don't support temperature, and max_output_tokens caps reasoning plus output, so a low cap can return an incomplete response instead of a shorter answer.
Deep dive
How it works
GPT-5 series reasoning models expose two separate controls:
- reasoning_effort controls how much the model thinks before it answers. Higher effort generally means more reasoning tokens.
- verbosity is a newer parameter that gives you more granular control over how concise the model's output is.
The requirement is shorter answers with the same amount of reasoning, so you leave reasoning_effort alone and set verbosity to low. Microsoft's sample request does exactly that: "text": {"verbosity": "low"}.
The other options, and when you'd use them
- reasoning_effort minimal: It changes how much the model thinks, which the requirement says to keep.
- max_output_tokens much lower: The cap covers reasoning, visible output and formatting tokens. If it's reached, the response can come back incomplete, possibly before any visible output, instead of shorter.
- Lower temperature: Reasoning models other than GPT-6 Astra don't support temperature.
Exam tip
Shorter output, same thinking = verbosity: low. Less thinking = reasoning_effort. max_output_tokens is a hard cap that can cut answers off.
References
Last reviewed: