OpenAI launches Decisions API for structured model evaluations

OpenAI released a new Decisions API this week, a dedicated endpoint designed to help models make structured judgments rather than generate open-ended text. The API, available at POST /v1/decisions, lets developers define an evaluation and get back a probability, a selection from predefined options, or a score against a rubric with ordered levels. It accepts both text and image inputs, so a team could feed it a delivery photo and ask whether a package is visible, or determine if a support ticket requires human escalation.

The move targets a persistent headache for anyone wiring an LLM into a workflow with real consequences. A false positive escalates something harmless. A false negative lets a genuine problem slip through unflagged. Developers have typically patched around this with manual classification rules, confidence thresholds, and separate evaluation pipelines bolted onto a general-purpose model. OpenAI's pitch is that the Decisions API handles that structure natively.

How the API works

Rather than using a general-purpose response for every task, developers define what the model needs to evaluate and the type of output required. The endpoint does not generate conversational text. It returns a clean answer based on the defined criteria. This narrow scope means the model is not being asked to reason broadly and then have its output parsed by downstream code - the structure is built into the call itself.

For teams already building agents for customer support, routing, or content validation, this removes the need to write classification scaffolding from scratch. The time saving is real. It also creates a direct dependency on OpenAI's infrastructure for every decision about escalating a complaint or validating a delivery photo.

What developers should watch

When the API's judgment calls disagree with a team's own standards, visibility into the reasoning becomes critical. Developers will want to know how much transparency the endpoint provides into why a decision landed where it did. Researchers should test whether this functions as a genuine primitive for agent reliability or simply wraps prompting patterns that teams already use. Independent testing on false positive and false negative rates will determine whether the API is safe for customer-facing applications.

For teams exploring structured AI outputs, related training on OpenAI API Courses and AI Agent Courses covers practical implementation patterns.

Why this matters for operations and support teams

If you manage a team that routes tickets, validates claims, or screens content, the Decisions API removes a layer of custom engineering you likely maintain today. That speeds up production timelines. The trade-off is that every binary call - escalate or don't, approve or flag - now runs through a third-party endpoint with defaults you do not control. Before adopting it, test the API against your own historical decisions and measure the gap. The time you save on scaffolding could cost you in edge cases you only discover after deployment.