Implement generative AI and agentic solutions 2.3 Optimize and operationalize generative AI systems
A reasoning model called through the Responses API sometimes returns a response with…
Study guide skill: Tune generation behavior, such as prompt engineering and adjusting model parameters
Question B8-025
A reasoning model called through the Responses API sometimes returns a response with status incomplete, incomplete_details.reason set to max_output_tokens, and no visible answer, although input and reasoning tokens are billed. What should you do?
Select one answer.
Answer: C
The cap set by max_output_tokens covers reasoning, visible output, and formatting tokens, and reaching it can return an incomplete response before any visible output, so reserve enough room and check status on every response. Reasoning models other than GPT-6 Astra don't support temperature or max_tokens. Higher reasoning effort generally produces more reasoning tokens, which makes the problem worse.
Deep dive
How it works
With reasoning models, the token cap covers more than the answer. max_output_tokens (Responses API) or max_completion_tokens (Chat Completions) covers reasoning tokens, visible output tokens and formatting tokens. Reasoning tokens are billed as output tokens.
If generation reaches the context window limit or your cap, the response comes back incomplete, with incomplete_details.reason set to max_output_tokens. Microsoft warns that this can happen before the model produces any visible output: you pay for input and reasoning tokens but get no answer. That's the symptom here.
The documented fixes:
- Reserve room. Reserve at least 25,000 tokens for reasoning and output while you're getting a feel for a workload, then tune the buffer once you know typical reasoning usage.
- Check status on every response, so your app handles this case instead of treating it as an empty result.
The other options, and when you'd use them
- Raise reasoning effort: Higher effort generally produces more reasoning tokens, which uses up the cap sooner.
- Lower temperature: Reasoning models other than GPT-6 Astra don't support temperature.
- Use max_tokens instead: max_tokens isn't supported on reasoning models other than GPT-6 Astra. Use max_output_tokens with the Responses API.
Exam tip
Reasoning model, status incomplete, no text = the cap was used up by reasoning. Raise max_output_tokens (start with 25,000 or more) and always check status.
References
Last reviewed: