OpenAI has launched the Decisions API in public beta. It turns textual content and pictures into typed solutions your code can department on. OpenAI group states the OpenAI Choices API runs about 10x quicker than the Responses API. It targets a standard sample: immediate an LLM, then parse its textual content right into a label.
TL;DR
- Dimension: GPT-6 Luna parameter depend shouldn’t be disclosed. Its model card lists a 1,050,000-token context window.
- Runs on: OpenAI-hosted API solely, by way of
POST /v1/choices. No open weights, no self-hosting. - Efficiency: About 10x quicker than the Responses API, per OpenAI.
- Finest: $0.10 per 1M enter tokens, with no output, cache-read or cache-write expenses.
- Backside line:
- Finest: quick, typed choices with possibilities.
- Worst: 1 mannequin, beta standing, no impartial evals but.
- Finest: quick, typed choices with possibilities.
What’s the OpenAI Choices API?
The Choices API is an OpenAI endpoint that evaluates textual content, photos or each and returns typed solutions. It doesn’t generate prose. A request has 3 fields: mannequin, enter and questions. The one supported mannequin as we speak is gpt-6-luna. OpenAI expects common availability in the coming weeks.
How does the Choices API work?
Every request carries shared proof plus an inventory of questions. Every query has a novel identify, a kind and directions. The response returns an solutions array keyed by these names.
What are the three query sorts?
- predicate: checks a situation and returns a
likelihoodfrom 0 to 1. Instance: does a product photograph present a crack, tear or dent? - selection: picks 1 worth from choices you provide. It additionally returns per-option
possibilitiesand aconfidencesubject. - rating: charges enter towards ordered
ranges, listed from 0. The rating is a probability-weighted common of degree indices.
OpenAI’s severity instance makes the maths concrete. Degree possibilities of 0.1, 0.7 and 0.2 yield a rating of 1.1. That worth sits between ‘Workaround obtainable’ and ‘Absolutely blocked’.
When do you have to use Structured Outputs as a substitute?
OpenAI attracts a transparent line. Use Choices for possibilities, decisions or scores. Use Structured Outputs to fill your personal JSON schema or write explanations. Use function calling when a mannequin should request a device name with arguments.
How briskly is it, and what’s the proof?
OpenAI’s doc claims about 10x quicker responses than the Responses API. DevDay protection put a call close to 150 ms, versus about 1.6 seconds for normal Luna calls. OpenAI has not revealed accuracy or calibration knowledge for the endpoint. The docs advise setting thresholds with labeled examples from your personal software.
window.addEventListener(‘message’,perform(e){
if(e.knowledge&&e.knowledge.mtpH){var f=doc.getElementById(‘mtp-dec-frame’);if(f&&e.supply===f.contentWindow){f.model.peak=e.knowledge.mtpH+’px’;}}
});
How a lot does it price, and the place are you able to deploy it?
With gpt-6-luna, enter prices $0.10 per 1M tokens. There are not any output-token, cache-read or cache-write expenses. Regional processing premiums and long-context multipliers nonetheless apply. Luna’s model card costs prompts above 272K tokens at 2x enter charges.
- Endpoint:
POST /v1/choiceson the OpenAI API, plus a Playground. - SDK minimums: Python 3.26.0, JavaScript 7.30.0, Go 3.73.0, Ruby 0.101.0, Java 4.78.0.
- Compliance: Zero Knowledge Retention and HIPAA for eligible prospects.
- Residency: United States and Europe (EEA plus Switzerland).
- Voice: choices can drive actions by way of client delegation with the Live API.
How does it examine with TypeSafe Jev?
The closest rival is TypeSafe Jev, launched September 15, 2026. Jev is a System One mannequin that returns typed values with calibrated possibilities. TypeSafe costs enter at $0.042 per 1M tokens, with free output. That makes OpenAI’s base charge about 2.4x larger. TypeSafe stories 70 to 500 ms end-to-end, measured from the US West Coast. Jev helps as much as 255 decisions however stays in early entry. OpenAI’s edge is picture enter, compliance choices and open beta entry.
| Function | OpenAI Decisions API | TypeSafe Jev 1.13 | GPT-6 Luna via Responses API |
|---|---|---|---|
| Launch | Beta, Oct 6, 2026 | Early entry, Sep 15, 2026 | GA, Sep 22, 2026 |
| Underlying mannequin | gpt-6-luna |
Jev (System One mannequin) | gpt-6-luna |
| Parameters | Not disclosed | Not disclosed | Not disclosed |
| Enter | Textual content, photos (inline base64) | Textual content and structured state | Textual content, photos |
| Output | Chance, selection or rating | Typed values with possibilities | Generated textual content (JSON by way of Structured Outputs) |
| Max choices per query | Not disclosed | As much as 255 | Not relevant |
| Pace (vendor-stated) | About 10x quicker than Responses API | 70 to 500 ms end-to-end | Baseline |
| Enter value per 1M tokens | $0.10 | $0.042 | $0.10 |
| Output value per 1M tokens | $0 | $0 | $0.50 |
| Compliance | ZDR, HIPAA (eligible); US and EU residency | Not disclosed | EU knowledge residency obtainable |
Key Takeaways
- OpenAI Choices API returns possibilities, decisions and scores, not prose.
- Enter prices $0.10 per 1M tokens; output is free.
- OpenAI claims about 10x quicker than the Responses API.
- TypeSafe Jev is cheaper at $0.042 per 1M enter tokens.
Try the Technical details here. All credit score goes to the researcher of this challenge. Additionally, be at liberty to comply with us on Twitter and don’t neglect to affix our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
The submit OpenAI Decisions API Hits Public Beta With 10x Faster Typed Answers appeared first on MarkTechPost.

