Function calling fails on “application/json” MIME type with the latest Gemini models

[ekiztk]

Describe the bug

It was working a few days ago. When calling the backend planner agent via Runner.run(), the SDK fails with:

• "Function calling with a response mime type: 'application/json' is unsupported"

This occurs even though the agent is explicitly returning a JSON-serializable Pydantic model (BackendPlannerAgentOutput).

Debug information

• Agents SDK version: 0.0.14
• Python version: 3.12.8

Repro steps

1. Install the Agents SDK and dependencies.

2. Create a file planner_test.py with the following minimal script:

   import json
   import asyncio
   from agents import Agent, Runner, AsyncOpenAI, OpenAIChatCompletionsModel
   from pydantic import BaseModel
   from enum import Enum
   from typing import List
   
   class FileActionEnum(str, Enum):
       CREATE = "create"
       UPDATE = "update"
       DELETE = "delete"
   
   class PIPManagerActionEnum(str, Enum):
       INSTALL = "install"
       UNINSTALL = "uninstall"
   
   class BackendFileTask(BaseModel):
       action: FileActionEnum
       file_path: str
       dependencies: List[str] = []
       description: str
   
   class BackendPackageTask(BaseModel):
       action: PIPManagerActionEnum
       package_name: str
   
   class BackendPlannerAgentOutput(BaseModel):
       file_tasks: List[BackendFileTask]
       package_tasks: List[BackendPackageTask]
       explanation: str
   
   async def run_backend_planner_agent():
       gemini_client = AsyncOpenAI(
           base_url="https://generativelanguage.googleapis.com/v1beta/openai/",
           api_key="YOUR_API_KEY"
       )
       payload = {
           "backend_stack_content": "dummy stack",
           "user_request": "Do something",
           "backend_project_directory": "."
       }
       agent = Agent(
           name="Planner",
           handoff_description="Test plan",
           instructions="Test prompt",
           model=OpenAIChatCompletionsModel(
               model="gemini-2.5-pro-preview-05-06",
               openai_client=gemini_client,
           ),
           tools=[],
           output_type=BackendPlannerAgentOutput
       )
       result = await Runner.run(agent, json.dumps(payload))
       return result.final_output
   
   if __name__ == "__main__":
       asyncio.run(run_backend_planner_agent())

[pakrym-oai]

This seems like a limitation of the gemini model (also mentioned in google-gemini/deprecated-generative-ai-python#515) it doesn't support both structured output and function calling at the same time.