OpenAI released the Decisions API in beta on October 6, 2026, and made it generally available on October 22: a new endpoint, POST /v1/decisions, that takes text or images plus a list of questions and returns typed answers (a probability, a choice or a score) instead of written text. It runs only on gpt-6-luna, charges $0.10 per million input tokens and nothing for output, and OpenAI says it “returns typed answers about 10x faster than the Responses API”. The product is a direct answer to Jev, the decision model from TypeSafe AI that spread through Vercel’s AI Gateway in September. This page reads OpenAI’s API changelog, the Decisions guide, the GPT-6 Luna model page and the pricing page, as of October 22, 2026.
On this page · 8 sectionsOpen
OpenAI released the Decisions API in public beta on October 6, 2026 and made it generally available on October 22: POST /v1/decisions takes text, images or JSON plus typed questions and returns answers instead of written text.
Three question types: predicate returns a probability that a condition is true, choice returns one option plus probabilities, score returns a weighted position on ordered levels.
It runs only on gpt-6-luna and charges $0.10 per million uncached input tokens with no output charge; since general availability, cache hits are discounted and cache writes are charged at Luna’s standard caching rates.
OpenAI says it returns typed answers about 10x faster than the Responses API; general availability on October 22, 2026 raised the limit to 400 questions per request, from 200.
TypeSafe’s Jev lists $0.042 per million input tokens, so 1M classifications of 500 tokens cost about $21 on Jev against $50 on OpenAI; neither vendor has published head-to-head accuracy.
Since general availability, images can be inline base64 or public HTTP or HTTPS URLs; file IDs are still refused. Zero Data Retention and HIPAA are available to eligible customers, with regional processing in the US and Europe.
§ 01What changed at general availability
On October 22, 2026, OpenAI’s API changelog marked the Decisions API generally available, and the guide now opens: “The Decisions API is now generally available.” Sixteen days after the beta, these are the changes.
| Item | Public beta, October 6 | General availability, October 22 |
|---|---|---|
| Questions per request | 200 | 400 |
| Images | Inline base64 only | Base64 or public HTTP or HTTPS URLs; images per question |
| JSON input | Not supported | input_json content parts |
| Prompt caching | No cache charges | Cache hits discounted, cache writes charged |
| Conversation history | Not listed | Messages plus function and tool calls and outputs |
| Usage reporting | Not listed | GET /v1/organization/usage/decisions |
| Model | gpt-6-luna | gpt-6-luna |
The pricing change is the one to read twice. Uncached input stays at $0.10 per million tokens and output stays free, but the guide now says: “Prompt cache hits receive discounted input pricing; cache writes are charged.” Those rates match Luna’s standard caching on the pricing page: $0.01 per million cached input tokens, and $0.125 per million for a 30-minute cache write or $0.25 for a 12-hour one. The cached prefix must be at least 1,024 tokens of shared input, so caching pays off when many questions run against the same document or image.
OpenAI’s launch post names five customers, Box, Datadog, Ramp, Sierra and Speak, and three internal users: the dots team, Personal Finance (categorizing transaction descriptions) and Communications. The one number a customer gives is Box’s: “In our evaluation, the Decisions API delivered ~5x faster route selection on 30 certificate of insurance cases, with the same accuracy as the Responses API.” That is 30 cases, a small sample, and OpenAI’s own headline claim stays “up to 10x faster”. The post also adds a native GPT-Live trigger, so a voice session can call Decisions as the user speaks, and more capacity in several Asian regions.
§ 02What OpenAI shipped
A request has three parts: the model (only gpt-6-luna today), the input (a text string, or user messages mixing text and images) and the questions, each with a unique name, a type and instructions. The response is an answers array keyed by those names. At launch, the guide said: “The Decisions API is in public beta, and we expect to GA in the coming weeks.” It went GA on October 22.
| Item | Decisions API |
|---|---|
| Released | October 6, 2026, public beta |
| General availability | October 22, 2026 |
| Endpoint | POST /v1/decisions |
| Model | gpt-6-luna only |
| Inputs | Text, images (inline base64 or public URLs) or JSON |
| Question types | Predicate, choice, score |
| Price | $0.10 per million uncached input tokens; no output charge; cache hits discounted, cache writes charged |
| Speed claim | About 10x faster than the Responses API |
| Data controls | Zero Data Retention and HIPAA for eligible customers; regional processing in the US and Europe |
| Playground | platform.openai.com/decisions |
§ 03The three question types
Each type returns a different kind of answer, and the guide gives one worked example per type: a damaged-product photo, a billing complaint routed to a department, and a bug scored for severity.
| Type | Use it to | What comes back |
|---|---|---|
| predicate | Check whether a condition holds, such as visible damage in a photo | A probability from 0 to 1 that the condition is true |
| choice | Pick one option from a fixed set, such as a department | The chosen value, a probability for every option, and a confidence |
| score | Rate an input on ordered levels, such as severity | The probability-weighted average of the level indices, which can fall between levels |
OpenAI’s advice on reading the numbers is to set thresholds from your own labeled examples, weighed by the cost of a false positive against a false negative. When the task is to produce an object of your own shape (extracted fields, a written explanation), the guide points back to Structured Outputs on the Responses API; Decisions is for answers, not generation.
§ 04What it costs, and the Jev comparison
The pricing is the headline. At beta the guide said: “You pay only for input tokens: there are no cache-read, cache-write, or output-token charges.” At general availability output stayed free and caching moved to Luna’s standard rates. Regional processing premiums and long-context multipliers still apply. The same model through the Responses API costs $0.10 per million input tokens and $0.50 per million output tokens, so a Decisions call removes the output bill entirely.
| Item | OpenAI Decisions API | TypeSafe Jev 1.13 | gpt-6-luna on Responses |
|---|---|---|---|
| Input price per 1M tokens | $0.10 | $0.042 | $0.10 |
| Output price per 1M tokens | None | Free | $0.50 |
| Images | Yes, base64 or public URLs | No, text only | Yes |
| Answer types | Predicate, choice, score | Boolean, choice, score | Free text or your JSON schema |
| Status | Generally available, October 22, 2026 | Early access; on Vercel AI Gateway | Generally available |
The worked example is our arithmetic: one million inputs of 500 tokens is 500 million tokens, $50 at OpenAI’s rate and $21 at Jev’s list price on our September 19 record. Jev stays cheaper per token and text-only; OpenAI’s version reads images, sits inside an account most teams already have, and carries ZDR and HIPAA eligibility. Neither vendor has published a head-to-head accuracy number, so the right test is your own labeled set on both.
§ 05What changes if you build on it
- Routing and triage get a native endpoint. Ticket routing, content moderation, lead scoring and relevance checks used to be a chat call parsed back into a label. Here the label and its probability come back typed.
- Images can be links now. Since general availability the endpoint takes public HTTP or HTTPS image URLs as well as base64 data URLs; uploaded file IDs are still refused.
- It is out of beta. OpenAI made it generally available on October 22, 2026, with caching, JSON input and 400 questions per request.
- Voice is wired in. The guide links a voice delegation pattern that pairs the Live API with Decisions to choose actions from spoken requests.
§ 06What we are watching
- More models. Whether models beyond
gpt-6-lunajoin the endpoint now that it is generally available. - Accuracy numbers. A published calibration or accuracy benchmark from OpenAI or an independent tester, against Jev or a Responses baseline.
- Jev’s reply. Any TypeSafe price or feature change after October 6.
- The 10x claim. Independent latency measurements of the speed gap OpenAI states.
§ 07Update log
- October 6, 2026: page opened at the public beta.
- October 22, 2026: general availability folded in: 400 questions per request, hosted image URLs, JSON input, and cache charges.
- October 22, 2026: OpenAI’s launch post added: named customers and Box’s 30-case speed test.
§ 08Sources
- OpenAI API changelog, October 6 and October 22, 2026: developers.openai.com
- OpenAI Decisions guide: developers.openai.com
- OpenAI, Build more responsive apps with the Decisions API, October 22, 2026: openai.com
- GPT-6 Luna model page: developers.openai.com
- OpenAI API pricing: developers.openai.com
- Our record of TypeSafe’s Jev, September 2026: cellcog.ai
Q1What is the OpenAI Decisions API?
The Decisions API is an OpenAI endpoint, POST /v1/decisions, released in public beta on October 6, 2026 and generally available since October 22. It takes text, images or JSON plus a list of questions and returns typed answers: a probability for a predicate, one option for a choice, or a score on ordered levels. It runs on gpt-6-luna.
Q2How much does the Decisions API cost?
With gpt-6-luna, uncached input costs $0.10 per million tokens and output is not charged, according to OpenAI’s guide. Since general availability, cache hits are discounted and cache writes are charged at Luna’s standard caching rates; regional processing premiums and long-context multipliers still apply.
Q3Is the Decisions API faster than the Responses API?
OpenAI says it returns typed answers about 10x faster than the Responses API. We have not seen an independent latency measurement yet.
Q4How does the Decisions API compare with TypeSafe's Jev?
Both return typed decisions with no output charge. Jev lists $0.042 per million input tokens and accepts text only; OpenAI charges $0.10 and accepts images. Neither has published a head-to-head accuracy benchmark, so test both on your own labeled data.
Q5Can the Decisions API read images?
Yes. Since general availability on October 22, 2026, images can be inline base64 data URLs or public HTTP or HTTPS URLs, and each question can carry its own images. File IDs are still not supported.


