# OpenAI Releases Decisions API: Structured Logic with Zero Output Cost

> OpenAI launched the Decisions API powered by gpt-6-luna. It executes bounded classification up to 10x faster than chat endpoints with zero output token fees.

- URL: https://roobrief.com/en/posts/openai-decisions-api
- Language: en
- Published: 2026-10-23T01:50:00.000Z
- Tags: ai, dev
- Source: https://developers.openai.com/api/docs/guides/decisions

**The Decisions API is a dedicated endpoint returning typed evaluations and classifications instead of generated text, running up to 10x faster on gpt-6-luna with zero output token fees.**

OpenAI announced on Oct. 22, 2026, the general availability of the **Decisions API** (`POST /v1/decisions`), an endpoint optimized for deterministic routing, evaluation, and classification tasks. Instead of prompting a conversational model to generate prose and parsing the result with regular expressions or JSON schemas, the API directly evaluates structured probabilities and discrete choices from input data.

## How It Differs from Chat Endpoints and What It Costs

While the standard Responses API is designed for open-ended conversational generation and code synthesis, the Decisions API focuses strictly on evaluating conditions to direct program execution.

| Attribute | Standard Responses API | Decisions API |
| --- | --- | --- |
| Primary Use | Prose, chat, and code generation | Predicate checks, classification, scoring |
| Supported Models | Diverse (GPT-4o, o-series, etc.) | Dedicated gpt-6-luna only |
| Execution Speed | Standard generation speed | Up to 10x faster (major p95 reduction) |
| Output Cost | Standard token pricing | $0 (returns structured data payload) |
| Input Cost | Varies by model | $0.10 per million input tokens |

The most significant architectural shift is that output tokens (a small piece of text, roughly part of a word, that AI models read and write) incur no charge. Because the endpoint returns bounded JSON arrays rather than streaming text character by character, developers pay only the base input fee of $0.10 per million tokens. Repetitive system instructions also qualify for prompt caching discounts.

## Three Question Types and Platform Integrations

Each API call accepts a shared input payload (text, image URLs, base64 payloads, or structured `input_json`) alongside an array of up to 400 distinct questions. The questions fall into three supported evaluation types:

- **Predicate:** Evaluates whether a condition is true or false, returning a probability value between 0 and 1 (e.g., assessing whether an uploaded image contains visible physical damage).
- **Choice:** Selects the most accurate category from a developer-defined array of choices, providing probabilities and confidence scores for each option (e.g., routing support tickets to the appropriate department).
- **Score:** Rates input data against a pre-configured rubric of two to ten ordered tiers.

The endpoint also integrates natively with GPT-Live for real-time voice and conversational systems. Applications can classify user intent or trigger avatar expression shifts mid-conversation without introducing noticeable latency.

## Best Use Cases and Current Constraints

The Decisions API is designed for latency-sensitive hot-path operations within software pipelines.

It serves as an efficient verification gate before autonomous agents trigger high-stakes actions, an automated triage filter for high-volume customer inquiries, and a real-time content moderation interceptor.

However, the endpoint cannot be used for tasks requiring open-ended explanations or conversational replies, as its output is strictly constrained to typed evaluation schemas. Furthermore, it only supports the lightweight gpt-6-luna model; complex coding logic and deep multi-step mathematical reasoning remain the domain of dedicated reasoning models.

## Comparison with Existing Alternatives

Developers previously solved high-speed routing and triage through different technical layers, each with clear trade-offs:

- **Embedding-based Semantic Routers:** Open-source tools like semantic-router match intents in local memory using vector cosine similarity. They execute in microseconds at zero network cost, but are limited to basic semantic proximity and cannot evaluate multi-variable rubrics or inspect images.
- **Dedicated Text Classifiers (such as Cohere Classify):** Traditional classification endpoints excel at categorized text tagging with few-shot training examples. However, they remain text-only and lack support for asking 400 heterogeneous questions across multimodal inputs in a single payload.
- **Standalone Decision Models (such as Jev):** Specialized non-generative classification models pioneered typed boolean and rubric evaluations, but OpenAI's offering differentiates itself through native multimodal image support, prompt caching, and direct integration into the GPT-Live audio pipeline.

## Sources

- [OpenAI Developers Guide — Decisions](https://developers.openai.com/api/docs/guides/decisions)
- [OpenAI News — Build more responsive apps with the Decisions API](https://openai.com/index/decisions-api/)