In October 2026 OpenAI released the Decisions API in public beta. Instead of writing text, it answers your questions with typed results: the probability that a condition is true, one option from a fixed list, or a score against levels you define. In my tests it answered in about a quarter of a second, roughly six times faster than the same model through the Responses API, and you only pay for input tokens. In this course you build TriageDesk, a support desk in Python that reads every new ticket, routes it to the right team, scores its severity, checks the attached screenshot, flags angry customers and decides the agent’s next action. Complex tickets go to a writing model that drafts a plan for a human to review.
Everything you see was recorded with the October 2026 public beta and checked against a real account. You will learn what the documentation does not tell you, measured with real calls:
- Predicates returned calibrated probabilities, while choice confidence changed by almost forty points just by reordering the options.
- How to pick thresholds from labeled tickets, as the official guide recommends.
- Real latency, real cost per ticket and a monthly estimate.
- How option descriptions become your support policy, and how overlapping descriptions steal decisions.
- How to protect your decisions with regression tests that call the real API.
How this course is made: I designed the project, wrote and ran every line of code against the real API, and reviewed every lesson. The narration uses an AI voice and the animations are generated from my scripts. The complete project, with thirty labeled tickets, a script per lesson, a test suite and a small JavaScript version, is included as a download.
