OpenAI Launches Decisions API in Public Beta

Dedicated endpoint trades conversational text for structured boolean, categorical, and scoring decisions in 150 milliseconds.

Autonomous agent pipelines spend substantial compute asking language models basic operational questions: should a tool run, is a permission valid, or which queue gets a ticket? Using standard chat endpoints for these checks burns seconds and bills developers for conversational filler nobody reads. On October 6, 2026, OpenAI addressed that friction by opening its Decisions API in public beta, deploying a specialized endpoint designed strictly for fast classification and routing.

What the Decisions API Actually Does

Hosted at POST /v1/decisions and powered exclusively by gpt-6-luna, the interface cuts out prose generation entirely. Instead of parsing messy text blocks, callers receive one of three rigid data types:

• Predicates (predicates): Returns a estimated probability between 0 and 1 that a given assertion is true.
• Choices (choices): Picks a single option from a developer-defined array, returning the index with a confidence score.
• Scores (scores): Rates input data against a numeric or ordinal rubric provided in the call.

OpenAI claims this stripped-down pipeline runs up to ten times faster than standard calls through its Responses API. In company documentation, text decisions post an average latency of roughly 150 milliseconds. Multimodal calls that evaluate images alongside text land around 0.8 seconds on a 3 Mbps connection.

Pricing, Context, and Minimum SDKs

The economic pitch is aggressive. OpenAI prices the service at $0.10 per million input tokens. Because the endpoint returns discrete logic rather than text streams, output tokens carry zero charge, and OpenAI applies no read or write fees for prompt cache.

The default window ingests up to 272,000 input tokens, alongside an extended tier supporting up to 1 million tokens for feeding large reference sets before making a call. Developers must upgrade to recent official SDK releases to access the endpoint: Python 3.26.0+, JavaScript 7.30.0+, Go 3.73.0+, Ruby 0.101.0+, or Java 4.78.0+.

Market Pressure and Current Limits

The release accelerates an emerging contest around micro-decision models. Sam Altman previewed the capability during OpenAI’s DevDay on September 29, 2026, trailing rival TypeSafe AI’s release of Jev, a specialized model tailored to multi-agent routing. Cloud competitors like Microsoft joined the bracket with Microsoft-Decision-1, creating a race toward lean, single-purpose inference.

Yet the beta exposes distinct boundaries. The endpoint remains locked to gpt-6-luna, with no support for lightweight distillations or multi-step reasoning models. Developer community feedback indicates that while single-turn probability calibration holds steady, reliability drops when chaining conditional decisions together. OpenAI expects general availability in the coming weeks, but hasn’t committed to a timeline for token caching or model flexibility.

Sources

• Decisions API Documentation — OpenAI (Official)
• Decisions API Public Beta Announcement — OpenAI Developer Community (Official)
• OpenAI API Changelog — OpenAI (Official)
• TechCrunch: OpenAI’s Jev clone — TechCrunch (Independent)
• StackOne: OpenAI Decisions API Analysis — StackOne (Independent)