BerriAI/litellm · error · AzureOpenAIError

Unexpected string response from Azure: {response[:500]}

Error message

Unexpected string response from Azure: {response[:500]}

What it means

After the sync Azure OpenAI chat completion request succeeds at the HTTP level, the SDK response must be a parsed model object, not a str. If it is a string, LiteLLM raises this 500 including the first 500 characters. Indicates an SDK/version mismatch or an exotic response shape from the gateway in front of Azure.

Source

Thrown at litellm/llms/azure/azure.py:350

                    api_version=api_version,
                    api_base=api_base,
                    api_key=api_key,
                    model=model,
                    client=client,
                    _is_async=False,
                    litellm_params=litellm_params,
                )
                if not isinstance(azure_client, (AzureOpenAI, OpenAI)):
                    raise AzureOpenAIError(
                        status_code=500,
                        message="azure_client is not an instance of AzureOpenAI or OpenAI",
                    )

                headers, response = self.make_sync_azure_openai_chat_completion_request(
                    azure_client=azure_client, data=data, timeout=timeout
                )
                if isinstance(response, str):
                    raise AzureOpenAIError(
                        status_code=500,
                        message=f"Unexpected string response from Azure: {response[:500]}",
                    )
                stringified_response: Final = response.model_dump()
                ## LOGGING
                logging_obj.post_call(
                    input=messages,
                    api_key=api_key,
                    original_response=stringified_response,
                    additional_args={
                        "headers": headers,
                        "api_version": api_version,
                        "api_base": api_base,
                    },
                )
                return convert_to_model_response_object(
                    response_object=stringified_response,
                    model_response_object=model_response,

View on GitHub (pinned to 6c2dcb801b)

Solutions

1. Align versions: pip install -U litellm openai so the openai package matches the version litellm declares.
2. Check the embedded response snippet in the message to see what the endpoint actually returned.
3. If a gateway (Cloudflare AI Gateway etc.) is in the path, verify it forwards Azure OpenAI responses unmodified.
4. Report with the snippet if it persists on matched versions.

Example fix

# before
pip install litellm==1.x openai==0.28

# after
pip install -U litellm openai  # let litellm pin a compatible openai

import litellm, openai
print('litellm', litellm.__version__, '| openai', openai.__version__)  # verify compatible pair before deploying

from litellm.exceptions import APIError

for attempt in range(2):
    try:
        resp = litellm.completion(model='azure/gpt-4o', messages=msgs)
        break
    except APIError as e:
        if 'Unexpected string response' not in str(e) or attempt == 1:
            raise
        time.sleep(1)

When it happens

Trigger: An incompatible openai package version returning raw strings from client.chat.completions.create; routing through a Cloudflare AI Gateway or proxy that returns non-standard payloads; monkeypatched SDK in tests.

Common situations: Upgrading litellm without upgrading (or with a downgraded) openai SDK; pip resolving an old openai version after another dependency pinned it.

Related errors

• azure_client is not an instance of AzureOpenAI or OpenAI
• Azure client is not an instance of AsyncAzureOpenAI or Async
• embedding_response is not an instance of EmbeddingResponse
• Unsupported provider config: {transcription_provider_config}
• Missing model or messages

AI-assisted analysis of BerriAI/litellm@6c2dcb801b (2026-08-15). Data as JSON: /api/errors/79b58168abe92e68. Report an issue: GitHub.