[BUG] Flow with Azure OpenAI models fails with "does not support response_format"

[gpeterlini]

Description

I'm trying to follow the Build your first flow guide from CrewAI with a gpt-5-mini deployment on Azure, but when I try running the flow with crewai flow kickoff I get the following error:

ValueError: The model azure/my-gpt-5-mini-deployment does not support response_format for provider 'azure'. Please remove response_format or use a supported model.

Returns the same error with azure/gpt-4.1-mini.
Checking if my deployments support response_format directly through LiteLLM returns True.

Steps to Reproduce

1. Follow Build your first flow guide
2. Set up Azure model according to the docs Azure configuration example with AZURE_API_KEY, AZURE_API_BASE and AZURE_API_VERSION keys on .env file
3. Edit LLM constructor on src/crewai_flow/main.py

# Initialize the LLM
llm = LLM(model="azure/my-gpt-5-mini-deployment", response_format=GuideOutline)

4. Edit crews/content_crew/config/agents.yaml and set "llm" to "azure/my-gpt-5-mini-deployment"
5. Run flow with crewai flow kickoff
6. Provide inputs, for example "python" and "beginner"

Expected behavior

The output described in the "Create your first flow" guide

Screenshots/Code snippets

Error with Azure:

Switching to a different provider like "gemini/gemini-2.0-flash", works fine:

Operating System

Other (specify in additional context)

Python Version

3.11

crewAI Version

0.165.1

crewAI Tools Version

--

Virtual Environment

Venv

Evidence

Checking my model deployment through LiteLLM directly seems to be fine.

from litellm import get_supported_openai_params

params = get_supported_openai_params(model="my-gpt-5-mini-deployment", custom_llm_provider="azure")

print("response_format" in params)
##Outputs "True"

Also according to Microsoft's docs, both gpt-4.1 and gpt-5 models support structured outputs.

Possible Solution

None

Additional context

OS: WSL - Debian 12

[sginji]

@gpeterlini do u see the same error if response_format is not present as well when you initialize the llm model?

[gpeterlini]

@gpeterlini do u see the same error if response_format is not present as well when you initialize the llm model?

No, the model will be initialized normally without response_format, but then the LLM will just output a string and not a GuideOutline pydantic object and anything that tries to access it's fields, like response.title, will throw an exception and cause the Flow to fail.

[ayushmangpta]

On Azure, it causes an issue where the deployment name contains dots or dashes.
This workaround solved it for me:

from crewai import LLM
import os

llm = LLM(
    model="azure/gpt-4.1",  # ⚠️ Use your exact deployment name
    api_key=os.environ["AZURE_API_KEY"],
    base_url="https://<URI>/openai/deployments/gpt-4.1/",
    api_version="version"
)

then in every agent i passed the llm

def name_agent(self) -> Agent:
    return Agent(
        config=<config>,
        verbose=True,
        tools=<tools>,
        llm=llm
    )