A returns workflow receives a photo of a cracked lamp. It needs one answer before routing the case: does the image show damage?
OpenAI's Decisions API is built for that kind of bounded question. The October 6 public beta uses gpt-6-luna with text and image inputs. Questions return predicates, choices from predefined options, or numeric scores. OpenAI reports answers up to 10× faster than using Luna through the Responses API; that is a provider claim, not a latency measurement from this site. API changelog.
Here is an illustrative request body for POST /v1/decisions. Replace the image placeholder with actual base64-encoded JPEG bytes:
{
"model": "gpt-6-luna",
"input": [{
"role": "user",
"content": [
{"type": "input_text", "text": "Inspect the lamp in this returns photo."},
{"type": "input_image", "image_url": "data:image/jpeg;base64,<JPEG_BYTES>"}
]
}],
"questions": [{
"type": "predicate",
"name": "visible_crack",
"instructions": "Is there a visible crack in the lamp? Ignore reflections and damage to the packaging."
}]
}Images must use inline data URLs. External image URLs and file IDs are not supported. Read the answer's type before its value: a predicate provides a probability, while a refusal provides no scored answer. Route refusals and uncertain cases to review rather than quietly interpreting them as “no damage.” Request reference.
The guide lists $0.10 per million input tokens, with no output, cache-read, or cache-write charge for this endpoint. Regional premiums and long-context multipliers apply. These are Decisions API rates; they are not general Luna prices for other endpoints.
Your application still chooses the next action. Test the classifier against labelled photos, including glare and ambiguous damage, before deciding which answers can be routed automatically.