Reading the kitchen panel.

Press Cmd+K (Mac) or Ctrl+K (Windows/Linux) on any assistant reply to open the kitchen panel. It slides in from the right and shows what actually happened for that specific turn.

The fields:

**MODEL.** Which model produced this reply (claude-sonnet-4-6 or claude-haiku-4-5). This is the model that actually served the response, not the model your composer was "set to" — the router may have downgraded to Haiku for budget reasons (see next field).

**ROUTING REASON.** One sentence explaining why the router chose that model. "User selected DEEP" means you explicitly asked for Sonnet. "Design mode (always DEEP)" means the chat is a design mode chat. "Design mode + scoped edit → FAST" means you had an element selected in the canvas, which routes to Haiku for sub-second edits. "Budget degrade → FAST" means you're near your monthly cap and got downgraded gracefully.

**TOKENS.** Input tokens (what we sent) and output tokens (what we generated). If prompt caching fired on this turn, you'll also see CACHE READ and CACHE WRITE — cache reads cost 10% of a normal input token, so a high CACHE READ number means low actual cost.

**TOOLS.** Whether web search fired and how many queries; whether extended thinking activated and how many thinking tokens.

**COST.** Approximate USD cost of this turn to us. Meaningful for heavy chats; usually negligible per turn.

The kitchen panel exists because the honest thing is to show you what happened. Most AI products hide the routing decision. Ours is one Cmd+K away.