gpt-5-mini Returns Empty `output_text` with Responses API

[PierrunoYT]

Confirm this is an issue with the Python library and not an underlying OpenAI API

• This is an issue with the Python library

Describe the bug

gpt-5-mini Returns Empty output_text with Responses API

Issue

When using gpt-5-mini with the Responses API, the returned output_text is frequently an empty string (""), even though the API call succeeds.

Expected Behavior

• response.output_text should contain the model's generated response.
• Should behave like gpt-4o-mini, which consistently returns text.

Actual Behavior

• response.output_text is empty.
• response.output contains only a ResponseReasoningItem of type "reasoning".
• No ResponseOutputMessage with actual content.

Reproduction Code

from openai import AsyncOpenAI

client = AsyncOpenAI()
response = await client.responses.create(
    model="gpt-5-mini",
    input=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Say hello in one sentence."}
    ],
    max_output_tokens=50,
)

print(f"output_text: '{response.output_text}'")  # Empty!
print(f"output length: {len(response.output)}")  # 1
print(f"output[0] type: {response.output[0].type}")  # reasoning


### To Reproduce

-

### Code snippets

```Python

OS

Windows

Python version

Python 3.11

Library version

Latest

[JavaJourneyer]

Hi @PierrunoYT , I am working on a PR to fix this , but in the meanwhile ,

User side workaround :

response = await client.responses.create(
model="gpt-5-mini",
input=[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Say hello in one sentence."},
],
modalities=["text"], # <- ensures a text message is produced
response_format={"type": "text"}, # <- alt way to force text
max_output_tokens=100,
)
print(response.output_text) # now non-empty

[PierrunoYT]

Hey

Thanks for the reply.
The strange thing is that its also happening with Rest API.

[JavaJourneyer]

@PierrunoYT :
output_text is working as designed — it concatenates only output_text blocks from message items. If the API response contains only a reasoning item and no message, output_text will be empty.

Why this happened
In the original case, the request hit a very low max_output_tokens limit before a message was produced, so the model returned only reasoning. With no message, there’s nothing for output_text to aggregate.

How to get output_text reliably
When calling gpt-5-mini with the Responses API:

Ask for text explicitly:

"text": { "format": { "type": "text" } }

Provide a high enough max_output_tokens so the model can produce the message after any reasoning.

Example REST request

curl https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5-mini",
    "input": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Say hello in one sentence."}
    ],
    "text": { "format": { "type": "text" } },
    "max_output_tokens": 200
  }'

This produces both a reasoning item and a message item, so output_text will contain the assistant’s reply.

If you also want the reasoning text
Parse it directly from the reasoning item in response.output.

[PierrunoYT]

So, is this a bug or normal behaviour? It's really confusing! I tried to create something with AI, but it switched back to another model because of this issue.

[JavaJourneyer]

Check after increasing the max tokens : "max_output_tokens": 200

You should get the message .

[JavaJourneyer]

I think its normal behavior

[PierrunoYT]

Look here

[PierrunoYT]

[JavaJourneyer]

Here you go :

  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5-mini",
    "input": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Say hello in one sentence."}
    ],
    "text": { "format": { "type": "text" } },
    "max_output_tokens": 200
  }'

output:

{
  "id": "resp_6899a601d9d08195aae5ce9c987c394b0096344ef7c9138d",
  "object": "response",
  "created_at": 1754899969,
  "status": "completed",
  "background": false,
  "error": null,
  "incomplete_details": null,
  "instructions": null,
  "max_output_tokens": 200,
  "max_tool_calls": null,
  "model": "gpt-5-mini-2025-08-07",
  "output": [
    {
      "id": "rs_6899a602da8c8195a4ad9c789ee1f8c00096344ef7c9138d",
      "type": "reasoning",
      "summary": []
    },
    {
      "id": "msg_6899a6057964819591366b3fadd76edc0096344ef7c9138d",
      "type": "message",
      "status": "completed",
      "content": [
        {
          "type": "output_text",
          "annotations": [],
          "logprobs": [],
          "text": "Hello, nice to meet you."
        }
      ],
      "role": "assistant"
    }
  ],
  "parallel_tool_calls": true,
  "previous_response_id": null,
  "prompt_cache_key": null,
  "reasoning": {
    "effort": "medium",
    "summary": null
  },
  "safety_identifier": null,
  "service_tier": "default",
  "store": true,
  "temperature": 1.0,
  "text": {
    "format": {
      "type": "text"
    },
    "verbosity": "medium"
  },
  "tool_choice": "auto",
  "tools": [],
  "top_logprobs": 0,
  "top_p": 1.0,
  "truncation": "disabled",
  "usage": {
    "input_tokens": 22,
    "input_tokens_details": {
      "cached_tokens": 0
    },
    "output_tokens": 141,
    "output_tokens_details": {
      "reasoning_tokens": 128
    },
    "total_tokens": 163
  },
  "user": null,
  "metadata": {}
}

[PierrunoYT]

Nice it works. Thx you.

I think it's still a bug tho.

[JavaJourneyer]

Now, I see that the max_tokens = 200 , you increased the output tokens and then it worked.

[PierrunoYT]

Yeah im not sure. The AI fixed it with your instructions.

It still thinks it's a bug in the library or whatever.

[BradleyVangelder]

Hi, I'm having the same issue with sending the API calls in JavaScript. If you check the OpenAI dashboard, you see that it's returning an empty output. Maybe it's a bug in the API itself?

[JavaJourneyer]

I am trying to put a patch in to make it behave like it used to with gpt-4-mini. Thanks for your patience. Also Its not a bug based on my understanding. you just need more output tokens. check the above comments