SIGNAL / TOOLS

OpenAI Decisions API: answer types, GPT-Live integration and performance limits

SOURCEOpenAI

PUBLISHEDOctober 22, 2026

LAYERTools

SIGNAL SCORE38/100

SXF VERIFIED BRIEF

OpenAI moved Decisions API to general availability, adding a bounded-output route for text, image and native JSON context, GPT-Live trigger integration, prompt caching and advertised US/EU inference residency. Developers define permitted answers instead of prompting a general text response and then parsing it.

WHY IT MATTERS

Routing a call, screening a command or scoring a document often needs a small, machine-readable decision rather than a paragraph. Predicate, choice and score outputs reduce answer ambiguity and can fit in latency-sensitive agent steps, provided confidence thresholds and human review are measured rather than assumed.

WHAT TO VERIFY

OpenAI reports up to 10x faster processing than GPT-6 Luna via Responses API, but that is a product claim and depends on workload and setup. The published customer examples have different baselines. Check the current pricing guide, regional access, error rates and calibration on the application's own dataset.

SXF RESEARCH EXPLAINER

What the technical evidence means.

Source-grounded analysis

TECHNICAL ANALYSIS / 01

How it differs from asking an LLM for JSON

A general text-generation call can be instructed to output a label, but downstream code still needs to parse and validate that generation. Decisions API asks for a bounded predicate, choice or score and returns structured results. A triage service could ask for the probability that a ticket is urgent, then choose one permitted department and score its severity without turning each step into free-text interpretation.

The three output types are not interchangeable: predicate is a calibrated-looking numerical estimate, choice is a selection among specified candidates, and score is the probability-weighted mean of ordered levels. A score is therefore a rubric-dependent ordinal summary, not a measurement with automatic real-world validity.

Check primary-source evidence ↗

TECHNICAL ANALYSIS / 02

Why batching and live voice matter

The October 22 announcement states that a single request can contain up to 400 questions and that GPT-Live sessions can attach decision triggers to conversation events. One example responds to transcript deltas with an allowed avatar expression and receives a decision.event. This is an architectural option for frequent lightweight classification within a running conversation.

At production scale, ask whether requests can share reusable context through prompt caching, whether a batch introduces undesirable tail latency, and whether a low-confidence result should stop an action. Teams should measure p50 and p95 latency, false escalations and per-decision total cost, not merely token price.

Check primary-source evidence ↗

TECHNICAL ANALYSIS / 03

Read the benchmark claims carefully

OpenAI advertises speed up to 10x relative to its GPT-6 Luna Responses API path. Box separately describes about 5x faster routing over 30 certificates of insurance at comparable accuracy; Sierra reports sub-280 ms p95 for its own integration. These are self-reported demonstrations from distinct tasks and cannot be combined into a universal SLA.

A good evaluation holds input examples, answer options, rubric and error tolerance constant when comparing Decisions with existing prompting. Human review remains essential for consequential decisions, particularly where the numerical result appears precise but the category boundary is subjective.

Check primary-source evidence ↗

PRIMARY-SOURCE FACTS

What the evidence establishes.

Reviewed 2026-10-31

Product availability

The announcement says Decisions API is generally available using GPT-6 Luna through a dedicated endpoint. Primary evidence ↗

Three answer types

predicate returns a probability from 0 to 1; choice selects a predefined option with probabilities/confidence; score averages ordered rubric levels. Primary evidence ↗

Batch and inputs

OpenAI reports text, image and native JSON context and support for up to 400 questions per request. Primary evidence ↗

Integration and residency

GA additions include native GPT-Live triggers, cached reusable context and inference residency in US and EU. Primary evidence ↗

Performance context

OpenAI states up to 10x faster than GPT-6 Luna via Responses API; Box reports about 5x faster selection on a 30-document routing test. Primary evidence ↗

LIMITATIONS

General availability is established by OpenAI's announcement, but customer testimonials and provider speed comparisons are not independently controlled benchmarks. Prices, request limits and residency eligibility may change; consult the current developer documentation.

TOPICSOpenAI

MODELSGPT-6 Luna GPT-6

OpenAI Decisions API: answer types, GPT-Live integration and performance limits

What the technical evidence means.

How it differs from asking an LLM for JSON

Why batching and live voice matter

Read the benchmark claims carefully

What the evidence establishes.

Keep the context connected.

Accelerating engineering for the physical world

Scaling Git at OpenAI

Introducing the Grid Protectors Program

LifeVal

Working with academic communities to build scientific understanding