✨ Feature Request: Allow Separate Models for Tool Execution and Final Response in OpenAI Agent SDK

Summary

Please add support in the OpenAI Agent SDK to split model usage between:

• One model for tool reasoning/execution (e.g., gpt-3.5-turbo)
• Another model for final response generation (e.g., o1)

Motivation

In many real-world agent workflows, the agent needs to:

1. Execute one or more tools (e.g., retrieve context, classify intent, generate a prompt)
2. Use the output to generate a high-quality final response

Using a single high-end model (like o1) for all steps increases cost unnecessarily.
Using a cheaper model (like gpt-3.5-turbo) for all steps reduces the final response quality.

Proposed API

Agent(
    name="Vector Assistant",
    reasoning_model="gpt-3.5-turbo",   # Handles all tool execution and logic
    response_model="o1",           # Used only once, after tools run
    ...
)