Empty Responses from GPT Reasoning Models

[Messatsu92]

Summary:
We are experiencing a recurring issue where requests to GPT reasoning models (versions 5.0, 5.1, and 5.2) consistently return empty responses across all our Azure subscriptions. This happens regardless of the deployment type (Data Zone Standard, Global Standard, etc.).

Details:

The issue occurs with reasoning level set to ‘medium’ or ‘high’.
The problem is reproducible via:
Our internal applications
The Foundry Playground
Postman
The subscription or deployment type does not affect the outcome.

Example Prompt:

As an expert in sustainability and fashion, identify and list the most impactful sustainability innovations in the fashion industry reported within the last 2 years. Provide a concise summary of each innovation, including its purpose, implementation, and impact. Present the output as a bullet-point list in plain text.
Example Response:

{
"choices": [
{
"content_filter_results": {},
"finish_reason": "length",
"index": 0,
"logprobs": null,
"message": {
"annotations": [],
"content": "",
"refusal": null,
"role": "assistant"
}
}
],
"created": 1770405547,
"id": "chatcmpl-D6Le7OgeWIKAJ8NUgfvRqQiQ18ZU6",
"model": "gpt-5-2025-08-07",
"object": "chat.completion",
"prompt_filter_results": [
{
"prompt_index": 0,
"content_filter_results": {
"hate": { "filtered": false, "severity": "safe" },
"self_harm": { "filtered": false, "severity": "safe" },
"sexual": { "filtered": false, "severity": "safe" },
"violence": { "filtered": false, "severity": "safe" }
}
}
],
"system_fingerprint": null,
"usage": {
"completion_tokens": 500,
"completion_tokens_details": {
"accepted_prediction_tokens": 0,
"audio_tokens": 0,
"reasoning_tokens": 500,
"rejected_prediction_tokens": 0
},
"prompt_tokens": 67,
"prompt_tokens_details": {
"audio_tokens": 0,
"cached_tokens": 0
},
"total_tokens": 567
}
}

Notes:

The content field of the assistant’s message is always empty.
The finish_reason is length, but no completion is generated.
No filter or refusal is triggered.

Request:
Could you please investigate this issue? Let us know if you need further details or specific logs.

[dantatepico]

Not sure if it's related but we're seeing issues with GPT-5 too. The SDK throws an error related to JSON, as though the response is empty.

What's strange is that if we look at the thread/conversation logs in Foundry, we can see that the responses are getting generated properly.

Azure/azure-sdk-for-net#55561

[leestott]

Hi @Messatsu92

This is a known and currently reproducible behavior with GPT‑5 reasoning models in Azure AI Foundry when reasoning effort is set to medium or high. In these cases, the model can consume the entire token budget on internal reasoning and emit only “reasoning” tokens, resulting in an empty assistant message even though the request succeeds and usage shows non‑zero reasoning tokens. The finish_reason is reported as length, but no message content is produced.

This has been observed across Foundry Playground, SDKs, REST (Postman), and multiple deployment types.

Why it looks confusing
• The request succeeds (no refusal, no content filter hit).
• usage.completion_tokens is non‑zero because reasoning tokens are counted.
• The assistant “content” is empty because no message output item was generated.
• Foundry conversation logs may show an internal response, but the API payload contains no message text.

This pattern is documented by Microsoft and OpenAI as expected behavior for GPT‑5 family reasoning models under certain configurations, not subscription- or region-specific failures.

Related and confirmed reports

• Microsoft Q&A: GPT‑5 / GPT‑5‑nano returning empty responses when reasoning is high and token budget is exhausted by reasoning: https://learn.microsoft.com/en-us/answers/questions/5590694/ai-foundry-model-gpt-5-nano-returns-empty-response

• Azure SDK for .NET issue showing downstream JSON parsing failures when GPT‑5 returns empty content blocks:
Azure/azure-sdk-for-net#55561

• OpenAI community threads explaining that GPT‑5 models may return only reasoning items and no message output, leaving content empty:
https://community.openai.com/t/gpt-5-responses-api-hundreds-of-calls-return-empty-completion-content-while-chat-works/1365210

Practical mitigations (short-term)
• Lower reasoning effort (e.g., low or minimal) for content-generation tasks.
• Increase max_completion_tokens / max_output_tokens so reasoning does not exhaust the budget before a message is emitted.
• Explicitly request text output (where supported) instead of relying on default aggregation.
• For pure text generation, consider non‑reasoning models (e.g., gpt‑4.1 / gpt‑4o family) until this behavior is improved.

Sorry, as of now, this is acknowledged behavior across GPT‑5 reasoning models in Microsoft Foundry and is under investigation/triage by Microsoft and SDK teams, as evidenced by the linked GitHub and Microsoft Q&A threads. @nicholasdbrady

[nicholasdbrady]

Acknowledged, thanks