Ai Foundry model GPT-5-nano returns Empty response from model.

Question

Ai Foundry model GPT-5-nano returns Empty response from model.

Luis TOBIAS 20

I have deployed a GPT-5-nano model on AI foundry. The model seems to go on and do long reasonings then return and empty response. The answer for a simple promtp like: hi this is a test, returns more than 800 tokens output for reasoning and an empty message output.

Is this a bug? Should i rather try to get the reasoning output, although i am not sure i should as it may be some unorgonized model "thoughts"

Sridhar M 1,220 Reputation points Microsoft External Staff Moderator

2025-10-21T08:16:33.9033333+00:00

Hi Luis TOBIAS,

Did you get any chance to review the above response.

Thank you!
Luis TOBIAS 20 Reputation points

2025-10-21T08:48:24.2233333+00:00

Hi Sridhar M,

I did get more useful answers by applying both recomendations: increasing the max_output_tokens and setting the reasoning effort to minimal. So far, i have not seen the "empty" text behavior.

Thanks,

Answer accepted by question author

0 additional answers

Your answer

Sridhar M 1,220 Reputation points Microsoft External Staff Moderator

2025-10-21T08:16:33.9033333+00:00

Hi Luis TOBIAS,

Did you get any chance to review the above response.

Thank you!
Luis TOBIAS 20 Reputation points

2025-10-21T08:48:24.2233333+00:00

Hi Sridhar M,

I did get more useful answers by applying both recomendations: increasing the max_output_tokens and setting the reasoning effort to minimal. So far, i have not seen the "empty" text behavior.

Thanks,

Answer 1

Hi Luis TOBIAS,

Behavior, not necessarily a bug: With the Responses API, GPT‑5‑nano/mini may emit only a reasoning item and no message item, so output_text is empty even when the call succeeds. This happens more often when token budgets are tight or reasoning effort is high. [github.com], [community.openai.com]

Wrong token parameter: Using max_tokens (legacy) with GPT‑5 models can cause invalid/ignored limits; use max_output_tokens (SDK) or max_completion_tokens (REST). [github.com]

High reasoning effort on “nano”: Default reasoning can consume hundreds of tokens before producing a message, sometimes exhausting the budget. For simple prompts, use minimal/low effort—or pick a non‑reasoning model for snappy chat. [free.blessedness.top], [platform.openai.com]

Safety & schema constraints: Azure’s content filters may strip completions (yielding empty content); structured outputs with invalid or incompatible schemas can also suppress responses. [free.blessedness.top], [free.blessedness.top], [free.blessedness.top]

Thank you!

Sridhar M 1,220 Reputation points Microsoft External Staff Moderator

2025-10-21T12:22:14.4833333+00:00

Hi Luis TOBIAS

Following up to see if the below answer was helpful. If this answers your query, do click Accept Answer and Yes for was this answer helpful.

Share via

Ai Foundry model GPT-5-nano returns Empty response from model.

0 additional answers

Your answer