xAI ships Grok 4.7 at Grok 4.6 prices, with split availability

If you are evaluating this for an agent loop, hold the channel constant. A result from grok-4.7-fast inside Cursor says nothing reliable about grok-4.7 on the API, and xAI prices the fast variant at twice the token rates, so a cost comparison across those two is not a comparison at all. xAI documents configurable reasoning effort at low, medium, high and xhigh, with high as the documented default, which makes effort a knob you own and a cost line you should log per run. The 200,000-token pricing boundary deserves a measurement rather than an estimate: instrument one representative agent transcript, count prompt tokens at each turn, and find out how often you cross it before you model spend.

The decision for a buyer is a governance one, and it has a clock on it. GitHub states that in Copilot Business and Enterprise, new models turn on automatically unless an administrator has already disabled the global default — so a third-party model can reach developer editors without anyone in your organisation choosing it. Check the current state of that policy rather than assuming it. Separately, xAI's documentation names US regional endpoints for the API; it does not establish coverage elsewhere, and an absent region is unknown, not global. If code or customer context would be sent to a new provider, route that through whatever review you already apply to processors, and record the answer before developers build habits around the model.