OpenAI's October 22 Decisions API announcement describes a specialized interface for bounded questions using GPT-6 Luna. Its supported outputs include probabilities, predefined choices, and scores. This illustrates a broader architecture choice: not every AI step needs to generate prose.
Use a bounded decision when the options are known
Routing a request, applying an ordered rubric, or deciding whether information is missing can fit a constrained output. The application can validate that the returned option is allowed before taking the next step.
However, format validity is not factual accuracy. A classifier that always returns an allowed value can still make costly mistakes.
Use generation for synthesis
Drafting an explanation, comparing evidence, or proposing a plan requires a different kind of output. These tasks benefit from source grounding, review, and a way to express uncertainty.
Combining the two can be useful: a bounded decision selects a safe route, then a generative model prepares a draft within that route. Do not let the generated explanation override authorization rules.
Test the boundary
Include ambiguous cases and options that should trigger manual review. Check probability calibration before using a confidence threshold to automate consequential actions.
Measure total workflow latency and cost rather than comparing isolated model calls. A fast classifier adds little value if it sends too many cases to the wrong downstream process.
The right distinction is not smarter versus simpler. It is the kind of evidence and output the application needs to make a controlled next step.
Sources & further reading
Reviewed October 26, 2026. Product availability and regulatory guidance can change. Vendor performance claims are not independent Rovio AI benchmarks.
Found this useful? Share it.