OpenAI's Decisions API is out of beta. Here is what it costs and how it compares to Jev.

OpenAI's Decisions API is now generally available. It turns GPT-6 Luna, OpenAI's smallest current model, into a decision model: you send text, images or JSON plus a list of questions, and it returns typed answers with probabilities instead of generated text. It costs $0.10 per 1M input tokens and OpenAI does not charge for output.

That puts OpenAI directly into a category that TypeSafe created with Jev and that Microsoft joined this month with Microsoft-Decision-1. If you run classification, routing or moderation steps through a full chat model today, this is the cheapest way to try the decision-model approach without leaving the OpenAI platform.

What OpenAI has confirmed

The details below come from OpenAI's Decisions guide and its API changelog.

• Timeline: the API entered public beta on October 6, 2026 and reached general availability on October 22, 2026.
• Model: requests go to POST /v1/decisions, and gpt-6-luna is the only supported model.
• Three question types: a predicate returns the probability that a condition is true, a choice picks one option from a list you supply and returns probabilities and a confidence value, and a score rates input against 2 to 10 ordered levels.
• Inputs: text, images (base64 or public URLs), JSON passed directly with input_json, and conversation history including tool calls. File IDs and audio input are not supported.
• Scale: up to 400 questions per request, double the 200 allowed during the beta.
• Speed: OpenAI says it returns typed answers about 10x faster than the Responses API.
• Pricing: $0.10 per 1M uncached input tokens and $0.01 per 1M cached input tokens, with no output-token charge. Regional processing premiums and long-context multipliers still apply.
• Compliance: Zero Data Retention, HIPAA support and regional processing in the US and EEA plus Switzerland.
• Voice: there is a native integration with GPT-Live, so a voice app can make decisions while the user is still speaking.

OpenAI is also clear about what the API is not for. If you need a generated object that follows a JSON schema, it points you to Structured Outputs, and for tool invocation it points you to function calling. Decisions only answers the questions you define.

What is reported

TechCrunch spoke to OpenAI's API product lead, Nikunj Handa, who said OpenAI first announced the product at its developer day in late September. He named speed as the main gain:

"Shrinking things from one to three seconds to 100, 120, 180 milliseconds." - Nikunj Handa, OpenAI, via TechCrunch

According to TechCrunch, OpenAI did not build a separate model for this. It reconfigured Luna to speed up inference and lower the cost. Early customers named in the report include Box, Datadog, Ramp and Speak, and Speak's CTO, Andrew Hsu, said the time between actions in its app dropped from 1.4 seconds to 0.13 seconds. These are customer and vendor figures, not independent measurements.

TechCrunch also reports that OpenAI's model is not the leader on public decision benchmarks. It places eighth on the AIM System One Index and fourteenth on the Artificial Analysis Decision Index, while the latest Jev model ranks seventh and third on the same two indices. TechCrunch adds that TypeSafe's models remain the most queried decision models on OpenRouter.

How it compares on price

• OpenAI Decisions (GPT-6 Luna): $0.10 per 1M input tokens, $0.01 cached, no output charge.
• TypeSafe Jev: $0.042 per 1M input tokens, output free.
• Microsoft-Decision-1: $0.042 per 1M input tokens, priced on input only, in public preview in Microsoft Foundry.
• GPT-6 Luna through the normal API: $0.10 per 1M input and $0.50 per 1M output tokens.

So OpenAI is more than twice the input price of Jev and Microsoft-Decision-1 on uncached input. Caching narrows that gap a lot: if you send the same long instructions or reference data on every call, the cached rate of $0.01 per 1M tokens is below both competitors' list prices. The pricing also makes the Decisions endpoint cheaper than calling Luna directly for the same task, since a classification answer no longer bills output tokens.

What this means for you

Decision models fit a narrow but common job: a step in an app or agent where the answer is one of a known set. Ticket triage, spam and moderation flags, routing a request to the right model or tool, checking whether an agent's output meets a rubric, and tagging images are all good matches. Open-ended writing, summarizing and code generation are not.

• If you already build on OpenAI: this is the easiest test. Your data controls, regional processing and billing stay the same, and you can move one classification step over and compare cost and latency against your current Luna or GPT-6.1 Sol calls.
• If raw price per call matters most: Jev and Microsoft-Decision-1 list lower uncached input prices, and Jev ranks higher on both benchmarks TechCrunch cites. Test them on your own labeled examples before committing.
• If you need voice or images: OpenAI's GPT-Live integration and image support in question instructions are the clearest differences from the other two.
• If you use AI through apps only: nothing changes in ChatGPT itself. You may notice faster routing and moderation inside third-party apps that adopt decision models.

Whichever you pick, measure on your own data. The probabilities these models return are only useful if they are calibrated for your categories, and the published benchmarks are early. A few hundred labeled examples from your real traffic will tell you more than any leaderboard.

Sources: OpenAI Decisions guide, OpenAI API changelog, TechCrunch.