Multi-turn customer-service tool use: look up a user, book or modify an order, and hold to the domain's policy across the conversation.
What this model claims, and what it has not measured yet.
Measured on real benchmark episodes against the Claude Fable 5 frontier anchor. Accuracy is the task-success gap in points; savings are per run, latency as p50 model time.
How far this model has come.
Everything that changes something, or spends anything, is behind the login.