OpenAI's Decisions API makes fast agent choices easier to test, not safe by default

OpenAI made its Decisions API generally available on 22 October 2026, positioning it as a specialised service for frequent, bounded choices inside applications and AI agents. The company says the GPT-6 Luna-powered endpoint can make these decisions up to ten times faster than running Luna through its general Responses API. The engineering implication is larger than latency: a constrained decision interface makes some model behaviour easier to measure, but does not turn a prediction into permission.

Applications repeatedly classify requests, choose the next tool or score whether an action needs review. Doing all of that through a conversational model adds tokens, orchestration delay and variability. Decisions exposes three result types: predicate for a probability between zero and one, choice for a selection among predefined answers, and score for an ordered rubric. Developers supply the question, allowed outputs and context as text, images or JSON.

Constrained answers reduce ambiguity, not error

OpenAI says a request can contain up to 400 questions, with caching and optional US or EU processing. The interface can also integrate with GPT-Live for decisions during voice sessions. These features make sense for classification and routing, where an application needs a small structured signal rather than another paragraph of reasoning.

However, a high reported probability should not be assumed to be calibrated for an organisation's own inputs. The same applies to a confidence field accompanying a choice. Before using outputs for security triage or customer eligibility, test the rates of false approval and false rejection against independently labelled data. Include out-of-distribution cases and changes in upstream data.

A model can propose the next action in milliseconds; a separate policy engine must decide whether the application may execute it.

Putting a fast judge in the hot path demands controls

OpenAI cites Datadog using Decisions to assess risk before agent-generated shell commands are executed, with the customer reporting much lower p95 decision latency than available frontier models. This is a vendor-published customer example, not a guarantee about arbitrary commands or deployments. A fast classifier may help prioritise review, but a permitted command still requires deterministic privilege, environment and change-control checks.

Teams should define a fail-closed outcome for missing responses, parse failures or high uncertainty when the action is consequential. Preserve the question version, relevant inputs, outcome and downstream execution result in audit logs. A test that verifies the classifier's JSON shape but not the resulting action is incomplete.

Optimise accepted outcomes rather than token price

The published price is $0.10 per million input tokens, with a lower rate for cached input and no output-token charge. OpenAI also reports customer performance examples such as Box's faster document routing and Sierra's lower-latency customer-interaction choices. Their workloads and baselines differ, so organisations should not combine those figures into a universal savings estimate.

Run the new endpoint in shadow mode alongside the existing classifier. Measure p50 and p95 latency, cost per accepted decision, human corrections, safety escapes and drift after changes. Release it gradually only when the measured outcomes meet preset thresholds. Structured decisions are valuable primitives; accountability still sits with the system that acts on them.

Produced through Devence Lab's automated research and publishing workflow, without prepublication human review. Devence Lab retains editorial responsibility. Sources are linked; report factual errors to contact@devencelab.com.

Written by the Devence Lab research team.