v2.45.0 breaking change: cache_write_tokens added as required field in InputTokensDetails breaks openai-agents SDK

[DMLoy]

Description

openai==2.45.0 added cache_write_tokens as a required field to InputTokensDetails in src/openai/types/responses/response_usage.py. This is a breaking change for openai-agents (and likely any other library that constructs InputTokensDetails directly) because openai-agents==0.18.0 creates that model with only cached_tokens:

# openai-agents/usage.py
self.input_tokens_details = InputTokensDetails(cached_tokens=self_cached + other_cached)

With openai==2.45.0, this immediately raises:

pydantic.ValidationError: 1 validation error for InputTokensDetails
cache_write_tokens
  Field required [type=missing, input_value={'cached_tokens': 0}, input_type=dict]
    For further information visit https://errors.pydantic.dev/2.13/v/missing

The error also surfaces when non-OpenAI providers (e.g. Google Gemini via the OpenAI-compatible endpoint) return a usage payload that only contains cached_tokens — the SDK's Pydantic model rejects it before it even reaches user code.

Steps to Reproduce

1. pip install openai==2.45.0 openai-agents==0.18.0
2. Run any agent against a non-OpenAI provider (e.g. Gemini via https://generativelanguage.googleapis.com/v1beta/openai/) — the provider returns input_tokens_details: {cached_tokens: 0} without cache_write_tokens.
3. Observe pydantic.ValidationError: cache_write_tokens Field required.

Or simply:

from openai.types.responses.response_usage import InputTokensDetails
InputTokensDetails(cached_tokens=0)  # raises ValidationError in 2.45.0

Diff

In v2.44.0 (source):

class InputTokensDetails(BaseModel):
    cached_tokens: int

In v2.45.0 (source):

class InputTokensDetails(BaseModel):
    cache_write_tokens: int   # ← NEW, required
    cached_tokens: int

Expected Behavior

cache_write_tokens should be Optional[int] = None (or at minimum have a default of 0) so that:

• Existing callers that construct InputTokensDetails(cached_tokens=...) don't break.
• Providers that don't return cache_write_tokens in their usage payload don't cause a validation error.

Workaround

Pin openai==2.44.0 until this is resolved.

Environment

• openai==2.45.0
• openai-agents==0.18.0
• Python 3.11

[DMLoy]

aligns with openai/openai-agents-python#3772