Default model is gpt-5.6-think-deeper. Attach text files or images (receipts, screenshots) — images are sent to Copilot's vision.
Token counts are estimates (~chars/4) — M365 Copilot does not expose exact token usage. Cost is what those tokens would have cost at standard pay-per-token list prices (M365 itself is flat-rate), so it is an estimate built on an estimate.
"From where" is the client IP (via X-Forwarded-For); geolocation isn't resolved (offline). "How" is derived from the User-Agent.
Each account is one set of M365 credentials with its own login, its own Copilot Studio agent and its own rate-limit budget. Pick one per chat, or pin an API token to it.
One token per client. Bind a token to an account and that client can only ever spend those credentials; leave it on Any and the client picks per request (X-M365-Account header or an account/model model id).
Point any OpenAI-compatible client here:
Default model: gpt-5.6-think-deeper. All models from the selector work here.