GPT-5.1/5.2: message.output_text.logprobs is empty when Structured Outputs (json_schema) is enabled in Responses API

API Bugs
,
1

Hi all,

I’m trying to use token logprobs as a lightweight “confidence” signal for classification labels. I can reliably get message.output_text.logprobs when Structured Outputs are NOT enabled, but when I enable Structured Outputs (json_schema), GPT-5.1 (and GPT-5.2 in my tests) returns logprobs=[] even though I request them via include=["message.output_text.logprobs"].

This seems related to other reports about missing logprobs / logprobs behavior:

What I’m doing

  • Endpoint: POST /v1/responses via openai-python (AsyncOpenAI)

  • Requesting logprobs: include=["message.output_text.logprobs"]

  • top_logprobs=1

  • For GPT-5.* I also set reasoning={"effort":"none"} (since logprobs are not supported for reasoning output)

Expected

If I request include=["message.output_text.logprobs"], I expect output[0].content[0].logprobs to contain token logprobs for the generated output text.

Actual

  • GPT-5.1 without structured outputs: logprobs present

  • GPT-5.1 with structured outputs (json_schema): logprobs=[]

  • GPT-4.1 with structured outputs (json_schema): logprobs present

Minimal reproducible example

from openai import AsyncOpenAI

async_client = AsyncOpenAI()

common = dict(
    model="gpt-5.1",
    input="Hello!",
    include=["message.output_text.logprobs"],
    top_logprobs=1,
    reasoning={"effort": "none"},
    max_output_tokens=64,
    temperature=1.0,
)

# no structured outputs
r1 = await async_client.responses.create(**common)

# structured outputs
r2 = await async_client.responses.create(
    **common,
    text={
        "format": {
            "type": "json_schema",
            "name": "HelloSchema",
            "schema": {
                "type": "object",
                "properties": {"x": {"type": "string"}},
                "required": ["x"],
                "additionalProperties": False,
            },
            "strict": True,
        },
        "verbosity": "low",
    },
)

print(f'{len(r1.output[0].content[0].logprobs)}, {len(r2.output[0].content[0].logprobs)}')

Output: 9, 0

1 Like