Decisions API
A decision-oriented API interface introduced by OpenAI, built on the Luna model and designed for developer scenarios that require AI-assisted decision-making.
Core Facts
OpenAI provides the open-source tool tiktoken, which allows developers to estimate token counts before calling the API.
The Decisions API focuses on decision-making scenarios, providing structured and actionable judgments given specific options or constraints.
The Decisions API runs on a model called GPT-6 Luna
Timeline (last 90 days)
The Decisions API supports visual input and can process multiple questions in parallel, but it does not have long-horizon reasoning capability, so complex tasks need to be handed to Astra or GPT-6.1 Sol.
The Decisions API does not train a new model; instead, based on the existing Luna weights, it compresses first-token latency through structured output, parallel inference, and inference stack optimization.
When the decision service is unavailable, the application must not simply collapse; fallback strategies need to be prepared: caching the most recent decision result, setting local fallback rules, or directly allowing non-critical decisions through.
The Decisions API is less suitable for core paths where decisions depend on massive real-time features and are extremely sensitive to single-call latency; such scenarios usually require moving computation locally.
Scenarios suitable for prioritizing the introduction of the Decisions API include: frequent but simple-rule judgments (such as content classification, permission checks, feature flags), product logic that requires consistent behavior across multiple clients, and teams whose rules iterate frequently and who do not want every change to go through the release process.
It is recommended to add dedicated instrumentation for decision calls, recording call latency, hit rate, and fallback counts, to verify whether the Decisions API truly brings response speed improvements.
In scenarios that require making judgments, traditional applications often call multiple services serially (checking rules, fetching configuration, running models, merging results), causing network round trips and accumulated latency.
OpenAI's Decisions API requires only 150 milliseconds per judgment and uses Luna, the smallest model in the GPT-6 family.
Starting from TypeSafe's release of Jev in mid-September 2026, at least eight companies launched AI decision models within a little over a month, including Cloudflare's Clef, the AWS Strands team's Decider, Perplexity's pplx-decider, Liquid AI's d1 series, OpenAI's Decisions API and Microsoft-Decision-1.
OpenAI officially claims the Decisions API is about 10 times faster than the Responses API, but it did not disclose benchmarks and did not publish head-to-head accuracy comparisons with competitors.
40 more timeline events
All Facts (20)
OpenAI provides the open-source tool tiktoken, which allows developers to estimate token counts before calling the API.
80%VerifiedThe Decisions API focuses on decision-making scenarios, providing structured and actionable judgments given specific options or constraints.
75%VerifiedThe Decisions API runs on a model called GPT-6 Luna
70%VerifiedDecision-1 can be called in three ways: the AI SDK's decision API, the OpenAI-compatible Decisions API, and the TypeSafe-compatible API.
65%VerifiedOpenAI announced that the Decisions API is open for public testing to all developers, powered by GPT-6 Luna and supporting text and image input.
70%VerifiedOpenAI launched a product called the Decisions API.
75%UnverifiedThe Decisions API does not train a new model; instead, based on the existing Luna weights, it compresses first-token latency through structured output, parallel inference, and inference stack optimization.
50%UnverifiedThe Decisions API supports visual input and can process multiple questions in parallel, but it does not have long-horizon reasoning capability, so complex tasks need to be handed to Astra or GPT-6.1 Sol.
50%UnverifiedThe Decisions API is less suitable for core paths where decisions depend on massive real-time features and are extremely sensitive to single-call latency; such scenarios usually require moving computation locally.
50%UnverifiedWhen the decision service is unavailable, the application must not simply collapse; fallback strategies need to be prepared: caching the most recent decision result, setting local fallback rules, or directly allowing non-critical decisions through.
50%UnverifiedIt is recommended to add dedicated instrumentation for decision calls, recording call latency, hit rate, and fallback counts, to verify whether the Decisions API truly brings response speed improvements.
50%UnverifiedIn scenarios that require making judgments, traditional applications often call multiple services serially (checking rules, fetching configuration, running models, merging results), causing network round trips and accumulated latency.
50%UnverifiedScenarios suitable for prioritizing the introduction of the Decisions API include: frequent but simple-rule judgments (such as content classification, permission checks, feature flags), product logic that requires consistent behavior across multiple clients, and teams whose rules iterate frequently and who do not want every change to go through the release process.
50%UnverifiedOpenAI's Decisions API requires only 150 milliseconds per judgment and uses Luna, the smallest model in the GPT-6 family.
50%UnverifiedOpenAI officially claims the Decisions API is about 10 times faster than the Responses API, but it did not disclose benchmarks and did not publish head-to-head accuracy comparisons with competitors.
50%UnverifiedThe OpenAI Decisions API offers zero data retention support and HIPAA compliance eligibility in the beta.
50%UnverifiedOpenAI Decisions API images must be passed in as inline base64, and hosted URLs and file IDs are both rejected.
50%UnverifiedOpenAI's official guidance draws three lines: the Decisions API for classification, routing, scoring, and gating that require a determination; Structured Outputs (Responses API) for scenarios that require extracting fields or written explanations in a structured form; and function calling for scenarios where the model triggers tools with parameters.
50%UnverifiedDecisions API output tokens are billed at $0, and billing is calculated on input only.
50%UnverifiedThe Decisions API currently supports only one model, gpt-6-luna, with no fallback model option.
50%