OpenAI’s Decisions API is fast and cheap—but what is it deciding?
The generally available service turns text, images or JSON into a probability, fixed choice or rubric score. Its low price could put model judgements inside routine workflows, while the evidence so far comes from OpenAI and early customers.
The short version
- What happened
- OpenAI says its Decisions API is now generally available and returns one of three bounded answer types: a probability, a choice from supplied options or a score against a rubric.
- Who it affects
- People may experience this service indirectly when software routes a support request, flags a document or chooses an agent’s next step. Faster automation can shorten waits, but a cheap misclassification can also be repeated at scale unless a person can review and correct it.
- What remains unproven
- OpenAI supplies both central sources; there is no independent benchmark or audit in the release.
Keep in mind
OpenAI supplies both central sources; there is no independent benchmark or audit in the release.
Customer examples use different tasks and disclose uneven denominators, so their speed, cost and accuracy claims cannot be compared as one study.
The ten-times-faster figure is an OpenAI comparison with GPT-6 Luna through its own Responses API, not a cross-provider result.
A bounded answer does not prove calibration, fairness or safety for any consequential use. Performance must be tested on the actual population, workflow and failure costs.
Source analysis · 2 linked sources. Primary analysed source date: 22 October 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 3 min The evidence, methods, implications and limitations
A smaller answer surface is the product
OpenAI made its Decisions API generally available on 22 October, offering developers a specialised route for repeated, bounded judgements such as classifying an input, choosing a next action or scoring against a rubric. Instead of writing a free-form response, the service returns a probability, one of the options supplied by the developer, or a probability-weighted score. It accepts text, images and native JSON, and can answer as many as 400 questions in one request.[1][2]
Cheap, fast calls can move models into the workflow’s hot path
The endpoint currently supports GPT-6 Luna and costs $0.10 per million input tokens, with cached input at $0.01 and no output-token fee. OpenAI says it is about ten times faster than using Luna through the Responses API. Early customers report faster routing, risk scoring and conversational actions, but their comparisons use different tasks and denominators. Box’s disclosed test covered 30 insurance-certificate cases; other examples do not publish a sample size or error distribution.[1][2]
The distinction that matters
Three answer types, three kinds of downstream control
The developer defines the question and what the application does with the returned value.
Predicate
Estimate whether a condition is true
Returns a probability from 0 to 1
Choice
Select from developer-supplied options
Returns a choice, probabilities and confidence
Score
Rate against ordered rubric levels
Returns a probability-weighted score
A confidence number is not a decision policy
Applications must still choose the question, allowed answers, threshold and action that follows. That matters when a score routes a customer, flags a payment, prioritises a worker’s case or lets an automated agent continue. The service’s narrow output can make integration easier, but it does not establish that the underlying judgement is fair, calibrated or safe for a particular population. Teams need representative tests, abstention and escalation rules, monitoring for drift and a way to reverse harmful outcomes.[1][2]
Watch the background, then check the evidence
Illustrative photo · not a video frame
The Royal Society · 2 min 31 sec · Background explainer
What is artificial intelligence? | The Royal Society
YouTube loads only when you choose; playback starts when you press play.
A short introduction to artificial intelligence from the Royal Society. Start here for basic concepts, then use current reporting to explore what has changed.
Evidence check: Published in 2018, before today’s generative AI tools. Useful for basic concepts, rather than an overview of current capabilities.
Published 2018-11-28. Source checked 14 October 2026. Check captions on the publisher’s video page.
Explore current AI technologyPhoto: ThisisEngineering / Unsplash · Licence
Continue the story
Read next
Continue exploring developer tools: compare the finding and its evidence limits.
Four agent projects are moving under one foundation—what changes for developers?
The Agentic AI Foundation has accepted projects from Anthropic, OpenAI, Hugging Face and Alibaba. Neutral governance may reduce vendor lock-in, but the announcement does not prove that implementations interoperate securely or reliably.
Read the next reportExplore the full source and evidence record
What this means for people
- People may experience this service indirectly when software routes a support request, flags a document or chooses an agent’s next step. Faster automation can shorten waits, but a cheap misclassification can also be repeated at scale unless a person can review and correct it.
Global context
OpenAI says regional processing is available in the United States and Europe, with Zero Data Retention and HIPAA use for eligible customers. Those options depend on eligibility and agreements; organisations elsewhere must check local data, consumer and automated-decision rules before deployment.
What the evidence does not yet show
- OpenAI supplies both central sources; there is no independent benchmark or audit in the release.
- Customer examples use different tasks and disclose uneven denominators, so their speed, cost and accuracy claims cannot be compared as one study.
- The ten-times-faster figure is an OpenAI comparison with GPT-6 Luna through its own Responses API, not a cross-provider result.
- A bounded answer does not prove calibration, fairness or safety for any consequential use. Performance must be tested on the actual population, workflow and failure costs.
What to watch next
- Independent, task-matched evaluations with published denominators and error distributions
- How developers set thresholds, abstention and human-review rules in consequential workflows
- Incident reporting, model updates and evidence that confidence remains calibrated over time
- Whether additional models, regions or auditable decision records become available
Living evidence record
Impact record IAI-0F15JAS
Evidence stage
Announced
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Not yet
Record status
Monitoring
Last checked
23 October 2026
Source trail
2 direct sources across 2 source types.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Related-source reporting disclosure
This record analyses 2 linked source records around the same underlying development. The extra records add method, date or context, but they do not by themselves constitute independent replication of every performance claim or predicted outcome.
Evidence trail
Sources used for this report
Links checked 23 October 2026
Research question, topics and regions
Topics: Developer tools · Automation · Model evaluation · API pricing · Operational risk
Regions: Global · United States · European Union · Asia
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.