GPT-6 Luna Decisions

GPT-6 Luna Decisions

Rubric scoring is only useful when the scores are consistent, and GPT-6 Luna Decisions was built for exactly that kind of consistency. The OpenAI decision model, served by Decisions API through the Jev playground, reads shared text, JSON, or images and returns typed answers with probabilities — and its Score answer type returns a value on an ordered scale you define, from 2 to 10 levels.

That structure turns subjective review into a measurable gate. Define a 3-level urgency scale, submit a complaint ticket, and GPT-6 Luna Decisions returns 2.8 out of 3 with the reasoning captured as data, not prose. Run the same rubric over a thousand tickets and the scores stay comparable, because every answer follows the same contract: one value, one scale, one probability attached. When a binary check is enough, the Noul type answers yes/no questions with a probability, and Choice handles rubrics that resolve to a named category with a distribution across up to 255 options.

The evaluation workflow is built in. The playground supports A/B rule comparison across up to 25 labeled cases, so you can tune level definitions against real examples before production. Requests accept up to 8 questions each, carry a 1,050,000-token context, and take up to 4 images per call — which covers screenshot-based quality scoring and document grading alongside plain text and JSON.

Costs are stable enough to run scoring continuously: $0.10 per million input tokens upstream, 1500 credits ($0.15) per million on the platform, one-credit minimum, 100 welcome credits for new accounts, and reservations released on failed calls.

For teams building evaluation harnesses, quality gates, or prioritization queues, GPT-6 Luna Decisions supplies the ordered scores and probabilities that make rubrics trustworthy at scale.