Inside the Master AI Gateway

The gateway is a router with a price list. Large architectural work goes to the strongest reasoning model; a one-line copy change goes to a small fast model. The routing decision is made from the size of the plan, not from the length of the prompt.

Caching is layered: a prefix cache for the stable part of the context, a cross-turn cache for search and web results, and a build cache for the bundler. Together they cut the cost of a typical follow-up turn by more than half.

Image generation runs through the same gateway, which is why you never paste an API key: the credit price is computed before the call and shown in the chat.