LLMs

OpenAI's Decisions API Goes GA: Structured Decision-Making at 10x Speed

OpenAI makes the Decisions API generally available, enabling bounded-answer inference 10x faster than GPT-6 Luna for classification and routing tasks.

Last verified:

Decisions API Reaches General Availability

OpenAI has made the Decisions API generally available as of October 22, 2026, according to the OpenAI Blog. Built on GPT-6 Luna and purpose-built for bounded-answer inference, the API delivers 10x lower latency than GPT-6 Luna through the Responses API for classification, routing, and scoring workflows. This release reflects OpenAI’s shift toward task-specific inference optimization: rather than forcing every decision through a general-purpose large language model, the Decisions API constrains Luna’s reasoning to a fixed set of possible answers, reducing compute and token overhead.

Three Answer Types for Workflow Automation

The API supports three structured response formats. Predicate answers return a probability estimate (0 to 1) for yes/no conditions—for example, whether a product image shows visible damage. Choice answers select one option from a developer-defined list and return both the selection and per-option confidence scores, useful for agent routing (continue, retry, escalate). Score answers apply a rubric, rating inputs against ordered severity or quality levels and returning a probability-weighted average.

According to the OpenAI Blog, internal teams are already deploying the API across diverse workflows: OpenAI’s Dots team reduces response times through faster decision-making; the Personal Finance team uses it to categorize transaction descriptions; the Communications team classifies media coverage by theme and narrative alignment. External customers are applying it to cost attribution and interactive language-learning experiences.

Expanded Capabilities and Regional Support

Since entering public beta, the Decisions API has doubled its per-request question limit to 400, added native JSON input support, integrated with GPT-Live for real-time voice decision-making, and enabled prompt caching for reused context. Inference residency is now available in the US and EU, with expanded capacity across several Asian regions. Image inputs are now supported directly in question instructions, reducing the preprocessing required for visual classification tasks.

Why This Matters

The Decisions API targets a historically expensive pain point: batch decision-making. Any workflow involving classification at scale—payment fraud detection, content moderation, customer routing, document triage—incurs latency and token costs when routed through a full language model. By constraining GPT-6 Luna to bounded outputs and optimizing inference specifically for that constraint, OpenAI reduces per-decision cost and latency, making agentic and real-time interactive applications more economically viable. Teams building high-frequency routing systems (e.g., customer support escalation, content categorization pipelines) should evaluate whether moving those decision nodes from Responses API to Decisions API improves cost or latency profiles. The native GPT-Live integration signals OpenAI’s bet that voice agents will require rapid, low-latency decision branching—a capability that general-purpose inference has historically struggled to provide at acceptable cost.

Frequently Asked Questions

What is the Decisions API and how is it different from other OpenAI APIs?

The Decisions API is optimized for bounded-answer tasks (classification, routing, scoring) on text, images, and JSON. Built on GPT-6 Luna, it achieves 10x lower latency than the Responses API by constraining inference to predefined answer types.

What answer types does the Decisions API support?

Three types: predicate (probability 0–1), choice (select from options with confidence scores), and score (probability-weighted rubric level).

How many decisions can I include in a single API request?

Up to 400 questions per request, doubled from the public beta limit of 200.

Is the Decisions API available globally?

Yes, with inference residency in the US and EU, and expanded capacity in several Asian regions.

#GPT-6 #API #inference #structured-outputs #latency