Auto in vscode uses models that I've disabled for my organization

[rschlack]

🏷️ Discussion Type

Question

💬 Feature/Topic Area

VS Code

Body

I have Claude Sonnet 4.6 disabled for my organization. This prevents me or my other developers from picking that model in VScode chat. However, if I have chat on Auto it will almost always use that model anyway. I can see the model being used in the Organization Settings AI Usage report. How can I get control of the models used if auto is ignoring the settings?

[dhruv-techdev]

I believe that this is a GitHub Copilot policy-enforcement bug.

GitHub’s documentation explicitly states that Auto model selection must follow the model policies configured by the organization or enterprise. Therefore, if Claude Sonnet 4.6 is disabled for your organization, Auto must not select it.

There is no additional VS Code setting required to enforce this. Disabling the model in:

Organization Settings → Copilot → Models

should remove it from both:

• Manual model selection
• Auto model selection

Since manual selection correctly blocks Claude Sonnet 4.6 but Auto continues using it, the organization policy is being applied inconsistently.

As a temporary workaround, developers should stop using Auto and manually select one of the organization’s approved models.

The organization owner should also open a GitHub Support ticket and include:

• Organization and enterprise names
• Screenshot showing Claude Sonnet 4.6 as disabled
• A recent AI Usage report entry showing that Auto used Claude Sonnet 4.6
• Exact date and time of the request
• Affected GitHub username
• VS Code version
• GitHub Copilot extension version

If the organization belongs to an enterprise, verify that no targeted enterprise model rule is enabling Claude Sonnet 4.6 for that organization. GitHub’s targeted model rules can control which models are available to specific organizations.

However, once the effective policy shows the model as disabled, Auto using that model is definitively incorrect and requires a fix from GitHub.

[initial-d]

The expected behavior is that Auto should not pick a model that is excluded by the effective org/enterprise policy.

GitHub’s Auto model selection docs say Auto chooses from supported models subject to policies, and specifically excludes models that are not available in the plan or are excluded by administrator policies:

https://docs.github.com/copilot/concepts/auto-model-selection

The model access docs also say that model access depends on plan, client, and whether the org/enterprise restricts specific models:

https://docs.github.com/en/copilot/how-tos/copilot-on-github/set-up-copilot/configure-access-to-ai-models

So if the effective policy for that user truly disables Claude Sonnet 4.6, but VS Code Auto still uses it, that looks like either a policy-resolution issue or a reporting/client mismatch.

I would check these before filing it as a product bug:

1. Confirm the user’s effective Copilot license source.

   If the user receives Copilot from more than one organization in the same enterprise, policy conflicts can matter. GitHub’s policy docs say that when the same policy is configured differently across organizations in the same enterprise, the least restrictive policy usually applies, with some exceptions:

   https://docs.github.com/en/copilot/concepts/policies

2. Check enterprise-level model settings, not only org-level settings.

   In Enterprise -> AI controls -> Copilot -> Configure allowed models, check whether Claude Sonnet 4.6 is:

   • enabled enterprise-wide
   • optional for organizations
   • allowed by a targeted model rule for this organization

   Targeted model rules can make the effective model list different from what you expect from only the org page.

3. Verify whether the AI Usage report row is showing the exact same user, organization, client, and timestamp as the VS Code session.

   If the user also has access through another org, GitHub.com, Copilot app, CLI, or another IDE, it is easy to attribute the model usage to the wrong policy context.

4. In VS Code, sign out/in or reload the Copilot extension after changing policies.

   I would not expect stale client state to permanently bypass a policy, but it is worth forcing a fresh entitlement/model list before collecting evidence.

5. Collect one clean repro packet:

   User:
   Organization:
   Enterprise:
   VS Code version:
   GitHub Copilot extension version:
   Policy screenshot showing Claude Sonnet 4.6 disabled:
   Any enterprise targeted model rules:
   AI Usage report row showing Auto -> Claude Sonnet 4.6:
   Timestamp/timezone:
   Prompt surface: VS Code Chat / Agent mode / Edit / other:
   

Temporary mitigation: ask developers to avoid Auto and manually select one of the allowed models until GitHub confirms the effective-policy behavior. But based on the Auto model selection docs, Auto using an explicitly disabled model is not something I would treat as intended behavior.