GPT-6 Luna Decisions

Shipping rules blindly is how pipelines drift, and GPT-6 Luna Decisions ships with the antidote: the Jev playground, where you can A/B test two rule sets against up to 25 labeled cases before a single production call goes out. This OpenAI decision model, hosted by Decisions API, reads shared text, JSON, or images and returns typed answers with probabilities — Choice for named options with distributions, Score for ordered scales, Noul for yes/no probability.

The playground workflow is refreshingly concrete. Capture rule set A and rule set B, attach your 25 labeled cases, and compare how each set behaves across every case side by side. Because answers are typed — an option name, a score between 2 and 10 levels, a yes/no probability — differences are unambiguous, and you can see exactly which rules misfire on which cases. What used to be a week of prompt tweaking becomes an afternoon of evidence.

When the winner is clear, the same contract moves to production through /v1/systemone, with answers read from data.result.answers. Up to 8 questions per call, a 1,050,000-token context, and multimodal input with up to 4 images mean the cases you tested are the payloads you will actually send.

Pricing is published rather than discovered: $0.10 per million input tokens upstream, 1500 credits ($0.15) per million on the platform, minimum billing of one credit, and 100 welcome credits for new accounts. Failed calls release their reservation, and the platform contract — 255 options accepted, 256 rejected — was verified against OpenRouter in October 2026.

Whether you are tuning routing rules, rubric thresholds, or moderation policy, GPT-6 Luna Decisions plus the Jev playground turns rule design from guesswork into measurement.