OpenAI officially launched its Decisions API on October 22, 2026, offering an interface dedicated to conditional logic, multiple-choice classification, and score calculation across text, image, and JSON inputs. Rather than generating descriptive, sentence-by-sentence explanations, the system returns closed choice signals and structured probability values. Running on OpenAIs flagship GPT-6 Luna model, the Decisions API slashes latency from the one to three seconds typical of the GPT-6 Luna Responses API down to approximately 100 to 180 milliseconds, increasing processing speed up to tenfold by delivering only predefined outcomes without narrative output.
The pricing and throughput structures also diverge sharply from traditional generation interfaces. OpenAI set pricing at $0.10 per million input tokens, dropping to $0.01 per million when utilizing prompt caching for frequently reused instructions or context. Because the system emits only choice or scoring signals, output tokens incur no charges at all. The production release doubles the single-request batch capacity from 200 questions during the public beta to 400 questions. The launch package additionally incorporates native GPT-Live integration along with inference data storage region settings covering the United States and Europe to address data sovereignty requirements.
Rather than training a distinct small model, OpenAI re-engineered the inference pathway of its general-purpose GPT-6 Luna model. Nikunj Handa, OpenAI API product lead, explained in an interview with TechCrunch that instead of building a specialized small model from scratch, the team restructured the broader general model around narrowed inference tasks, noting that colleagues internally discovered numerous applications right away. The strategic move positions OpenAI directly against single-purpose decision models such as TypeSafes Zev through serving-layer efficiencies. For software engineers building agentic workflows, the setup cuts post-processing parsing routines and mitigates inter-agent communication delays, though architecting safeguards for unanticipated inputs outside predefined choices remains an ongoing engineering priority.