Unrestricted caching keyed by generated types causes memory leak in multi-threaded regimes

[rona-sh]

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

In the OpenAI.responses.parse function, as part of validating the formatted response so it matches the required text_format (denoted hereinafter MyClass), the code creates the type ParsedResponseOutputMessage[MyClass], which is later passed to an unbounded lru_cache of pydantic.TypeAdapter.

In a multi-threaded setting, pydantic generates the type anew, and so its hash changes in every run, increasing the cache size ad infinitum. Concretely, a standard webserver that uses the responses.parse function is affected.

This issue reproduces under any model, and regardless of user input or target class.

To Reproduce

Consider the following snippet:

from openai import OpenAI
from pydantic import BaseModel
import psutil


class Fact(BaseModel):
    fact: str


client = OpenAI()
model = "gpt-4.1-nano"

def f():
    _ = client.responses.parse(model=model, input="Give a fun fact", text_format=Fact)
    print(psutil.Process().memory_info().rss / 2**20)

f invokes a responses.parse call, and prints the memory usage (in MB).

When running it on a single thread, the memory usage changes minimally:

for _ in range(10):
    f()

However, when running it on multiple threads (such as in default webserver), more memory is used for every request:

import time
import threading

for _ in range(10):
    time.sleep(0.1)
    threading.Thread(target=f).start()

Code snippets

OS

macOS

Python version

Python v3.12.11

Library version

openai v1.107.0