How Initial Model Selection Affects Subagents in GitHub Copilot

What you can’t find in VS Code GitHub Copilot documentation is a clear explanation of the following:

💡 If you expect that subagents will be used to complete your prompt, then for the prompt use the most expensive model among all models that you expect to be used in the entire multi-agent workflow.

Subagents are restricted from using more expensive models than the initial model in a user prompt.

If your initial prompt uses a 1x model, any subagent will use the model the subagent is configured to use only if this model is <= 1x model (1x, 0.33x, 0x). If the subagent is configured to use a more expensive model (3x, 30x), it will use the initial 1x model instead.

If your initial prompt uses a 3x model, any subagent will use the model the subagent is configured to use only if this model is <= 3x model (3x, 1x, 0.33x, 0x). If a subagent is configured to use a more expensive model (30x), it will use the initial 3x model instead.

You can’t use GPT-5.4 (1x model) for an orchestrator and Claude 4.6 (3x model) for its subagents. The subagents will use GPT-5.4 instead.

You pay only for prompts initiated by a user. You don’t pay when an agent calls a subagent. Therefore, this behavior (it was different in the past) is intended to prevent abuse – the ability of cheaper and free models to initiate a workflow that uses more expensive models.