All tools / OpenAI Decisions API
Source-checked updateOpenAI Decisions API: When to Use It and What GA Changes
OpenAI made the Decisions API generally available on October 22, 2026. It turns supplied evidence into a probability, a predefined choice or a rubric score. Use it for a bounded application decision, such as routing a support request. Official launch announcement.
Sources checked October 23, 2026 · No hands-on product test
What changed at general availability?
The release adds native JSON input, images inside individual question instructions, prompt caching, GPT-Live integration and up to 400 questions per request. The dedicated endpoint is POST /v1/decisions, currently using gpt-6-luna. Current developer guide.
Choose the right answer type
Predicate
A probability for a condition. Example: does this request contain enough information to continue?
Choice
One allowed option. Example: route a message to billing, technical support or manual triage.
Score
A rating against defined criteria. Example: assign an urgency level using a written support rubric.
Decisions or Responses?
Choose Decisions when the permitted answers are known before the request. Choose Responses with Structured Outputs when you need an object following your own schema, an extracted record or a written explanation. A decision result does not execute an action: your application controls what happens next. Input and output boundaries.
For a routing experiment, define an explicit manual-review option rather than forcing every unfamiliar message into a confident-looking category. Evaluate false positives and false negatives separately; a lower response time does not establish better decisions.
Pricing and speed: compare the right things
OpenAI lists uncached input at $0.10 per million tokens and cached input at $0.01 per million, with no output-token charges. Cache writes have separate rates; regional processing and long-context conditions can affect billing. Launch pricing · Current pricing and availability
The advertised speed gain is up to 10× against GPT-6 Luna through Responses. This is a vendor comparison, not a latency promise for your network or workload. We have not run a paid API test.
Before connecting it to live actions
- Choose one repeated decision and define its allowed outcomes.
- Build a small evaluation set containing ambiguous and out-of-scope inputs, with human-reviewed expected answers.
- Measure agreement, failure handling, latency and billed input for the same examples.
- Keep consequential actions behind your existing checks and permissions. Roll out only after the evaluation answers the business question.