A fast routing answer is useful only when the application can handle the wrong destination. Define the cost of a mistake before setting a decision threshold.
A dedicated interface for bounded decisions
OpenAI’s October 22 announcement makes the Decisions API generally available with GPT-6 Luna. The endpoint evaluates supplied material and returns typed answers to defined questions.
The current guide distinguishes three question types. A predicate estimates whether a condition is true. A choice selects from supplied options. A score evaluates ordered levels and returns a probability-weighted result.
These outputs can support routing, classification, and prioritization. The application still decides what each answer permits it to do.
What changed since the beta
The guide raises the request limit from 200 to 400 questions. It also describes native JSON input, image support within questions, prompt caching, and integration with GPT-Live.
OpenAI describes input-only billing for the dedicated Decisions endpoint. A shared input and its questions still consume tokens; inspect actual usage and applicable cache charges before estimating total cost.
The company reports substantial latency improvements over the Responses API. That comparison depends on its task and test conditions. We have not reproduced the result.
Measure the mistakes that affect the workflow
Start with a fixed set of previously resolved support requests. Include clear examples, ambiguous requests, and cases that belong outside the available categories.
- Keep the expected destination and reason for each example.
- Record the selected category and returned probabilities.
- Count incorrect automatic routes separately from manual escalations.
- Measure latency and actual billed input usage.
- Review the cases that crossed an automatic-action threshold incorrectly.
A returned probability does not establish calibration on your data. For example, group predictions within the same probability range and compare them with known outcomes before using the value as a risk estimate.
Give uncertain cases a defined destination
For a low-impact support queue, a mistaken route may cost staff time. A permission decision or payment action can have a much larger consequence. Use different acceptance rules for those workflows.
An escalation route needs an owner and a deadline. Otherwise, reducing incorrect automatic decisions can move work into an unattended queue.
Compare the full process with the existing rule or model baseline. Include accepted outcomes, correction time, and total cost across retries, rather than reporting endpoint speed alone.
Our Jev article covers another typed decision system. The agent cost-control guide connects routing choices with actual spending.
Sources
Which decision in your application can safely proceed automatically, and where should uncertainty send it?
Checked October 25, 2026. Prepared with AI-assisted research and drafting, with claims checked against the linked sources. MustHave.ai has not performed the proposed product tests.
