should we preserve schemas in additionalProperties? is that supported by openai?

Yes, preserving structured additionalProperties schemas is supported by OpenAI's API. The Structured Outputs documentation shows that Dict[str, T] types work correctly because they use structured additionalProperties like {"additionalProperties": {"type": "number"}}.

The fix only sets additionalProperties=False when it's True (from Pydantic extra="allow") or missing. For Dict[str, T] fields, Pydantic generates structured schemas which we now preserve.

The original code was breaking Dict[str, Decimal] by forcing False everywhere, which would reject all dynamic keys.

Yes, preserving structured additionalProperties schemas is supported by OpenAI's API. The Structured Outputs documentation shows that Dict[str, T] types work correctly because they use structured additionalProperties like {"additionalProperties": {"type": "number"}}.

Interestingly, the documentation you linked to doesn't say that, as far as I can tell. All of the examples are "additionalProperties": false. However, that does seem to be correct in terms of what the API actually does.

In the hopes of speeding along the deployment of this fix, below are some examples demonstrating @yashwantbezawada's comment to be correct. The second one could be adapted into a test case, and suggests that adding "properties": {} when absent might be convenient if any users use this kind of schema.

# uv run --with 'ipython,pydantic,openai @ git+https://github.com/yashwantbezawada/openai-python.git@bcf94fbe44a49e09df0a1624809bdb46d17e6156' ipython
import os
from typing import Dict
from openai import OpenAI
from openai.lib import _pydantic
from pydantic import BaseModel, ConfigDict, RootModel

# Ensure you have your API key set
# os.environ["OPENAI_API_KEY"] = "sk-..." 

client = OpenAI()

# Direct json schema:
response = client.chat.completions.create(
    model="gpt-5-mini",
    messages=[
        {"role": "user", "content": "Generate a price map for 'apple', 'banana', and 'cherry'."}
    ],
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "price_map",
            "strict": False, 
            "schema": {
                "type": "object",
                # The fix: You must explicitly provide 'properties', even if empty.
                "properties": {},  
                "additionalProperties": {
                    "type": "number"
                }
            }
        }
    }
)

print(response.choices[0].message.content)


class PriceMap(RootModel[Dict[str, int]]):
    root: Dict[str, int]
    model_config = ConfigDict(
        json_schema_extra={
            "properties": {} # Force this key to exist
        }
    )


assert PriceMap.model_json_schema()["additionalProperties"] == {"type": "integer"}
print(_pydantic.to_strict_json_schema(PriceMap))
response = client.beta.chat.completions.parse(
    model="gpt-5-mini",
    messages=[
        {"role": "user", "content": "Generate a price map for 'apple', 'banana', and 'cherry'."}
    ],
    response_format=PriceMap,
)

print(response.choices[0].message.parsed)

That outputs:

{"apple":1.20,"banana":0.50,"cherry":3.00}
{'additionalProperties': {'type': 'integer'}, 'properties': {}, 'title': 'PriceMap', 'type': 'object', 'required': []}
root={'apple': 0, 'banana': 0, 'cherry': 0}

Note the second example also succeeds on the latest published version of openai python library (uv run --with 'ipython,pydantic,openai==2.9.0' ipython), outputting:

{"apple":0.99,"banana":0.29,"cherry":3.49}
{'additionalProperties': {'type': 'integer'}, 'properties': {}, 'title': 'PriceMap', 'type': 'object', 'required': []}
root={'apple': 150, 'banana': 0, 'cherry': 3}

However on commit cdebce9 (uv run --with 'ipython,pydantic,openai @ git+https://github.com/yashwantbezawada/openai-python.git@cdebce941252c8817447108b395b90997d45ec40' ipython), the request succeeds, but silently outputs incorrect values:

{"apple":0.75,"banana":0.30,"cherry":3.50}
{'additionalProperties': False, 'properties': {}, 'title': 'PriceMap', 'type': 'object', 'required': []}
root={}

So failing to preserve schemas in additionalProperties would be a breaking change, and make the JSON vs Pydantic APIs less consistent.

This PR seems like a simple fix for an unambiguous bug. My comment above gave quite precise testing instructions demonstrating that it works and that the old versions failed, in addition to the test case added in the PR. Is there anything I can do to help with getting this PR merged?

This issue is making it harder for my employer (Akasa) to continue using the OpenAI API. We're currently using a fork of the client library from 3 months ago which complicates our python dependency setup and makes debugging issues with use of your APIs harder. Lower probability, but it's also a potential security/compliance issue if any critical bugs are found in in the client library.

Not to put too fine a point on it, but Anthropic recently launched structured outputs on AWS bedrock, covered by the AWS BAA. This has made it feasible for us to switch to another vendor, though no decision has been made as of yet. Their client library does not have this bug, and integrating with it was quite straightforward.