What changed
longer effective context, better tool-call structure under load, updated cache behavior, pricing adjustments on output tokens.
What it means for tool use
fewer retries on complex tool chains. Better adherence to typed tool schemas. Cheaper to leave cache hints in the system prompt.
What we changed
flipped our concierge demo to Opus 4.6. Kept Haiku 4.5 on routing calls where latency matters more than reasoning depth. Logged the switch in our agent-runs receipts.