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"

Anonymous

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

Anonymous

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

Anonymous

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. [learn.microsoft.com], [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. [learn.microsoft.com], [learn.microsoft.com], [learn.microsoft.com]

Thank you!

Anonymous

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.
Pat Garvey 0 Reputation points

2026-01-05T20:53:16.45+00:00

What good is the model-router then? I thought it should return the "best" model with a cursory look at the prompt. If it always picks a nano model, and more than VERY simple reasoning is needed, the api call will fail with no content. (Tokens > 1000)

Share via

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

0 additional answers

Your answer