Azure AI Foundry Responses API returns 401 with Entra ID despite resource-scoped inference role

Question

Azure AI Foundry Responses API returns 401 with Entra ID despite resource-scoped inference role

Maurice Fluelling 0

I am testing a non-production Azure AI Foundry deployment using U.S. Data Zone Standard. Authentication uses Microsoft Entra ID; no API keys or desktop-embedded credentials are permitted.

Token acquisition succeeds. The evaluation identity was temporarily assigned Cognitive Services OpenAI User at the exact Foundry resource scope. After waiting more than five minutes for propagation, a single synthetic request to the Responses API returned HTTP 401 PermissionDenied before inference. No model tokens were consumed.

Could Microsoft clarify:

The currently supported Responses API endpoint and Entra token audience.
The least-privilege RBAC role and required assignment scope.
Whether the role must include Microsoft.CognitiveServices/accounts/OpenAI/responses/write or a newer equivalent.
Whether any Foundry project, provider, deployment, or network configuration is additionally required.
Whether a current gpt-5.4-mini U.S. Data Zone Standard deployment supports this authentication path.

The temporary role was removed after testing. No customer or production data was transmitted.

Sai Kiran Mudavath 185 Reputation points Microsoft External Staff Moderator

2026-10-10T04:51:33.0833333+00:00

Hi @Maurice Fluelling ,

Thank you for sharing the details of your testing. I reviewed the Microsoft documentation for the Azure OpenAI Responses API and Microsoft Entra ID authentication. Based on your findings, I recommend verifying the following configuration.

1. Verify the endpoint and token scope

For direct Azure OpenAI inference, use:

https://<resource-name>.openai.azure.com/openai/v1/responses

Acquire the Microsoft Entra ID access token using the scope:

https://ai.azure.com/.default

Ensure that the endpoint corresponds to the intended Azure OpenAI resource and that the request uses the configured deployment name.

Reference: Use the Azure OpenAI Responses API

2. Verify role assignment

The built-in Cognitive Services OpenAI User role supports Microsoft Entra-authenticated inference through this endpoint.

Please confirm that the role is assigned to the identity actually making the request, at the appropriate Azure OpenAI resource scope. Since you removed the role after testing, restore it temporarily for a controlled retest.

Reference: Azure built-in roles for AI and machine learning

3. Check the resource configuration

Please verify that:

The resource has a custom subdomain configured.

The requested model deployment exists on the resource represented by the endpoint.

Network access restrictions permit the request from your calling environment.

The token belongs to the intended tenant and identity.

For direct inference through the Azure OpenAI v1 endpoint, a separate Foundry project is not a prerequisite.

Reference: Authentication and authorization in Microsoft Foundry

4. Confirm model availability

Microsoft's documentation lists gpt-5.4-mini, version 2026-03-17, for the Responses API and Data Zone Standard deployment in the documented U.S. regions, including Central US, East US, East US 2, North Central US, South Central US, West US, and West US 3.

Please confirm that your deployment uses the documented model version and is available in the resource and region you are targeting.

Reference: Region availability for Foundry Models sold by Azure

5. If the 401 persists

Since you have already allowed time for role propagation, please recheck the endpoint, token scope, calling identity, and role assignment.

If the issue persists, please share the sanitized HTTP error response, the approximate UTC time of a failed request, the resource region, and any available request correlation ID. Please do not share access tokens, API keys, client secrets, or other credentials.

These details can help narrow down the remaining authentication or authorization issue.

I hope this helps clarify the supported configuration and the next troubleshooting steps.

Thank You!

1 answer

Your answer

Answer 1

For an Azure OpenAI gpt-5.4-mini deployment, call https://<resource>.openai.azure.com/openai/v1/responses with a token for https://ai.azure.com/.default, and assign Cognitive Services OpenAI User to the calling identity at that Azure OpenAI/Foundry resource scope.

Correct endpoint and token audience

Use:

POST https://<resource>.openai.azure.com/openai/v1/responses
Authorization: Bearer <token>
Content-Type: application/json

Acquire the token with this OAuth scope:

https://ai.azure.com/.default

The current Responses API documentation uses the Azure OpenAI v1 endpoint—not the project endpoint under *.services.ai.azure.com/api/projects/...—for direct model inference. The request’s model value must be the deployment name, which may differ from the underlying model name.^1^

A minimal Python test is:

from azure.identity import DefaultAzureCredential, get_bearer_token_provider
from openai import OpenAI

token_provider = get_bearer_token_provider(
    DefaultAzureCredential(), "https://ai.azure.com/.default"
)
client = OpenAI(
    base_url="https://<resource>.openai.azure.com/openai/v1/",
    api_key=token_provider,
)
result = client.responses.create(
    model="<deployment-name>", input="Return exactly: OK"
)
print(result.output_text)

This endpoint-and-token pattern is the documented keyless Responses API configuration.^2^

Least-privilege RBAC for this route

For direct Azure OpenAI inference, Cognitive Services OpenAI User is the appropriate built-in inference role. Assign it to the actual calling user, service principal, or managed identity at the resource represented by <resource>.openai.azure.com. That role permits Entra-authenticated inference calls.^3^

The current built-in definition includes:

Microsoft.CognitiveServices/accounts/OpenAI/responses/*

Therefore, a separate Microsoft.CognitiveServices/accounts/OpenAI/responses/write permission is not the documented operation name, and a custom role is unnecessary when Cognitive Services OpenAI User is used.^4^

After temporarily restoring the role, verify that it is attached to the same principal whose oid is in the token:

az role assignment list \
  --assignee <principal-object-id> \
  --scope <foundry-resource-id> \
  --include-inherited \
  --output table

Since the previous test already allowed more than five minutes, propagation alone is not a useful next diagnosis.

Do not mix the two inference surfaces

Microsoft’s Foundry Models keyless-authentication guidance also documents a different authorization pattern: Cognitive Services User at the Foundry resource, with Microsoft.CognitiveServices/accounts/MaaS/* for a custom role.^5^

That applies to the Foundry Models inference surface. It should not be substituted for the Azure OpenAI v1 Responses route above. In particular:

Azure OpenAI route: *.openai.azure.com/openai/v1/responses
Azure OpenAI role family: accounts/OpenAI/...
Foundry Models inference role family: accounts/MaaS/...
Foundry project endpoint: *.services.ai.azure.com/api/projects/<project-name>, used for project SDK/agent operations rather than the direct Azure OpenAI model route.^6^

A route/role-family mismatch is therefore the next item to eliminate.

Additional prerequisites

For the direct Responses API call:

A deployed Azure OpenAI model is required.
Microsoft Entra ID is supported and recommended.
The resource must have a custom subdomain for token-based authentication.^7^
No additional Foundry project or project managed identity is documented as necessary for direct model inference.
Use the deployment name in model; a wrong deployment name normally leads to a not-found response rather than establishing an RBAC failure.^1^
If public network access is disabled or Private Link is configured, the caller must use whatever network path the resource configuration permits; the supplied evidence does not establish whether this resource has such restrictions.

`gpt-5.4-mini` U.S. Data Zone Standard support

gpt-5.4-mini, version 2026-03-17, is listed as supported by the Responses API.^1^ It is also listed for Data Zone Standard deployment in the documented U.S. regions: centralus, eastus, eastus2, northcentralus, southcentralus, westus, and westus3.^8^

Thus, U.S. Data Zone Standard does not itself preclude this Entra authentication path. If the canonical endpoint, audience, deployment name, principal, and resource-scoped role all match and the request still returns 401 PermissionDenied, capture the response apim-request-id, timestamp, resource region, endpoint host, token aud/oid claims, and role-assignment output for an Azure support case.

References

Azure AI Foundry Responses API returns 401 with Entra ID despite resource-scoped inference role

1. Verify the endpoint and token scope

2. Verify role assignment

3. Check the resource configuration

4. Confirm model availability

5. If the 401 persists

1 answer

Correct endpoint and token audience

Least-privilege RBAC for this route

Do not mix the two inference surfaces

Additional prerequisites

gpt-5.4-mini U.S. Data Zone Standard support

Your answer

`gpt-5.4-mini` U.S. Data Zone Standard support