OpenAI Releases GPT-6 Luna to Power Decisions API
OpenAI has launched its new Decisions API powered by the GPT-6 Luna model, giving developers a streamlined way to route requests and classify data without complex output-parsing code.
At DevDay 2026, OpenAI introduced the Decisions API, which is powered by a new specialized model variant called GPT-6 Luna. Designed specifically for classification, routing, and agent-control tasks, Luna allows developers to provide a question, a set of allowed answers, and text or image context. The model then selects a single answer from the predefined list. This launch marks the introduction of Luna as a distinct member of the GPT-6 family, alongside other task-specialized variants like Astra and Sol.
For developers, the Decisions API simplifies the integration of classification tasks. Traditionally, using general-purpose models for routing required complex prompt engineering, structured-output schemas, and custom parsing logic to handle invalid responses. By embedding the allowed answer set directly into the API contract, Luna eliminates the need for extra validation and retry code. Applications can use the model to route support tickets to specific teams, assign moderation labels, classify documents like receipts, or select the next tool for an autonomous agent.
Alongside Luna, OpenAI announced GPT-6.1 Sol, an upgraded model tailored for agentic coding, computer use, and professional tasks. Sol delivers performance close to Astra but at one-fifth of Astra's standard input and output token prices. OpenAI also introduced an Ultrafast tier that achieves generation speeds of up to 300 tokens per second in Codex, representing an eightfold speed increase for Codex and a sixfold increase through the standard API.
The Decisions API is currently available in a limited preview for selected API customers, with a broader rollout planned for the coming days. While OpenAI has not yet shared pricing, rate limits, latency metrics, or choice limits for Luna, developers can begin testing the preview by shadowing existing classifiers. This allows teams to build confusion matrices and evaluate how Luna handles ambiguous inputs or adversarial prompts before deploying it to production.
This is our own summary of reporting by AlphaSignal
