← Back to feed

OpenAI made the Decisions API generally available on October 22, resolving a class of latency and cost problem that affects nearly every production agent or application: the thousands of small, bounded decisions — classify this, route that, score this output — that accumulate into bottlenecks at scale.

What It Does

The Decisions API takes GPT-6 Luna and constrains it to questions with defined answer sets. Developers specify a question and its allowed answers; the API returns one of three output types:

• Predicate: probability (0 to 1) that a condition is true
• Choice: one of the caller’s supplied values, with probabilities and confidence
• Score: probability-weighted average of rubric level indices

Input can be text, images, or native JSON. Up to 400 questions per request, double the previous public beta limit.

The speed claim is up to 10x faster than GPT-6 Luna through the Responses API. That gap is architectural: Luna through Responses is a general-purpose API call with full token generation overhead. Decisions constrains the output space and optimizes accordingly.

Pricing

• $0.10 per million input tokens
• $0.01 per million input tokens for cached context
• No output-token charges

The pricing model is unusual for OpenAI. Most API calls charge for both input and output. Decisions charges only for input, which makes the cost predictable for high-volume classification and routing workloads where output is a fixed token or number.

Inference residency is available in the US and EU.

What Was Added Since Public Beta

The GA build includes several additions beyond the initial beta:

• Native GPT-Live integration: routes real-time voice conversations through Decisions for interruption handling and next-action selection
• Prompt caching: reused context at the $0.01/M rate
• Inference residency: US and EU regional processing for compliance use cases
• Native JSON inputs: structured data without serialization to text
• Images in question instructions: vision-capable decision calls
• 400 questions per request: previously 200

The GPT-Live integration is the most architecturally interesting. Live voice applications need sub-second decisions about when to interrupt, what to say next, and whether to hand off. Routing those decisions through a dedicated API rather than a general LLM call addresses a real latency constraint.

Where Teams Are Using It

OpenAI’s own deployment gives a picture of the use cases. The company’s Personal Finance team uses it to categorize transaction descriptions; its Communications team uses it to classify press coverage by theme and score whether coverage supports or challenges specific narratives.

Customers are using it for AI cost attribution and language-learning interactivity, according to OpenAI.

The pattern is consistent: replacing an LLM call that answers a multi-token question with a constrained API call that returns a probability or a token from a defined set. The quality bar is the same model; the overhead is lower.

Context

GPT-6 Luna launched in September 2026 as OpenAI’s fastest, lowest-cost general model at $0.10/$0.50 per million input/output tokens. The Decisions API uses the same model but removes output-token overhead entirely for bounded decision tasks, making it cheaper than the base Luna API for anything that fits the predicate/choice/score pattern.

Zero Data Retention is available to eligible customers.