OpenAIModel fails with OpenRouter Gemini: missing "choices" and "created" keys

[gkeb]

Initial Checks

• I confirm that I'm using the latest version of Pydantic AI
• I confirm that I searched for my issue in https://github.com/pydantic/pydantic-ai/issues before opening this issue

Description

Summary

When using OpenAIModel with OpenRouter and the model google/gemini-2.0-flash-exp:free, the agent fails with:

TypeError: 'NoneType' object is not subscriptable

This happens because the response returned by OpenRouter for Gemini does not include the choices or created keys expected by _process_response().

---

Steps to Reproduce

from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.providers.openai import OpenAIProvider

model = OpenAIModel(
    'google/gemini-2.0-flash-exp:free',
    provider=OpenAIProvider(
        base_url='https://openrouter.ai/api/v1',
        api_key='sk-or-v1-...',
    ),
)


agent = Agent(model=model, system_prompt="Be helpful.")
result = agent.run_sync("Tell me a joke.")

Observed Error

• response.created is None, causing datetime.fromtimestamp() to fail
• response.choices is None, causing choices[0] to throw

Suggestion

• Add a fallback timestamp = datetime.now() when created is missing
• Handle cases where choices is missing — or validate the response format before attempting to parse
• Possibly support custom response parsers for models that aren't 100% OpenAI-compatible

Why it matters

OpenRouter is a growing ecosystem for multimodel access, and many of the available models (Gemini, Cohere, etc.) do not fully follow the OpenAI ChatCompletion schema. Supporting them would improve compatibility.

Let me know if you'd like me to submit a PR or share sample responses.

Thanks for the awesome tool!

Example Code

Python, Pydantic AI & LLM client version

Python 3.12
pydantic>=2.11.4
pydantic-ai-slim[openai]>=0.2.4
pydantic-settings>=2.9.1

[wdhorton]

I was able to reproduce this. It only happens sometimes--since OpenRouter has two providers for this model, I'm guessing it happens for one of the providers but not the other.

[gkeb]

Probably yes. If I set up OpenAI or DeepSeek it works. The problem practically only occurs with Google models. In any case, I only noticed it with them.

[Kludex]

I can't reproduce it. Also, trying many times is hard because there's a lot of 429.

[Kludex]

I've created the OpenRouterProvider to make it easier to work with it, also, you can now use Agent('openrouter:<model_name>') to use OpenRouter - Assuming you have the OPENROUTER_API_KEY set.

[Kludex]

See this: https://github.com/pydantic/pydantic-ai/pull/1778/files#r2097604759.

It didn't fail... How can I reproduce it?

[TechNickAI]

A workaround until #1870 is properly implemented:

OpenRouter sometimes returns malformed responses where response.created or response.choices is None, causing pydantic-ai to crash with TypeError: 'NoneType' object cannot be interpreted as an integer.

This custom model class patches the timestamp issue and properly triggers fallback models when OpenRouter returns invalid responses:

from dataclasses import dataclass
from datetime import UTC, datetime

from pydantic_ai.exceptions import ModelHTTPError
from pydantic_ai.messages import ModelResponse
from pydantic_ai.models.openai import OpenAIModel, chat


@dataclass(init=False)
class OpenRouterModel(OpenAIModel):
    """
    Fixes OpenRouter's missing timestamps and malformed responses.
    Raises proper errors to trigger fallback models when needed.
    """

    def _process_response(self, response: chat.ChatCompletion) -> ModelResponse:
        """Handle OpenRouter's quirks before processing."""
        # Check for completely missing response
        if response is None:
            raise ModelHTTPError(
                status_code=502,
                model_name=self.model_name,
                body={"error": "Received None response from OpenRouter"},
            )

        # Fix missing timestamp
        if not hasattr(response, "created") or response.created is None:
            response.created = int(datetime.now(UTC).timestamp())

        # Check for missing choices array
        if not hasattr(response, "choices") or response.choices is None:
            raise ModelHTTPError(
                status_code=502,
                model_name=self.model_name,
                body={"error": "OpenRouter returned response with no choices"},
            )

        return super()._process_response(response)


# Usage:
model = OpenRouterModel(
    model_name="google/gemini-2.0-flash-exp:free",
    base_url="https://openrouter.ai/api/v1",
    api_key="your-key-here",
)

This handles both issues and ensures proper fallback behavior when using FallbackModel.

[Kludex]

This works as expected:

from pydantic_ai import Agent

agent = Agent('openrouter:google/gemini-2.5-pro', system_prompt='Be helpful.')
result = agent.run_sync('Tell me a joke.')
print(result)