OpenAI's Decisions API now generally available
OpenAI made the Decisions API generally available on October 22, 2026, targeting agent developers who need fast, low-cost decision-making at scale OpenAI. The API is purpose-built for bounded-answer selections—picking among predefined options—rather than generating full text responses, a design shift that fundamentally changes how agents can operate in production.
Built for speed and cost efficiency
The Decisions API leverages GPT-6 Luna through a dedicated endpoint and accepts text, images, and structured inputs. OpenAI claims the API is up to 10 times faster than GPT-6 Luna through the standard Responses API for these use cases OpenAI, a material performance gain for agent workflows that must make thousands of decisions per second.
Pricing is aggressive: $0.10 per million input tokens, with cached input tokens costing $0.01 per million tokens. Critically, OpenAI charges no output-token fees—a structural advantage for high-volume decision workflows where repetitive selections would otherwise drive costs OpenAI.
Regional availability and data retention
The API is available for regional processing in the United States and European Union. For eligible customers, zero-data-retention access is also available, addressing compliance concerns in regulated industries where data must not persist on OpenAI infrastructure OpenAI.
What this means for agent deployment
The Decisions API addresses a structural bottleneck in the agent economy: the cost and latency of turning language-model reasoning into action. Agents that route customer requests, flag content, approve transactions, or rank search results no longer need to pay for full-text generation or wait for token streaming. Instead, they get a binary classification, selection from a menu, or structured choice in milliseconds.
This architecture also reduces hallucination risk. Agents operating within bounded answer spaces cannot generate false outputs—the model selects from what you provide. For compliance-heavy use cases (lending decisions, fraud detection, access control), this constraint is a feature, not a limitation.
The release reflects a broader industry shift: as agents move from research demonstrations into production systems, the infrastructure is specializing. General-purpose text generation is no longer the bottleneck. Speed, cost, and reliability for specific agent tasks—decisions, tool calling, structured extraction—are.
