AIAI NO. ---

OpenAI Decisions API now generally available on GPT-6 Luna

OpenAI has opened its fast decision model to all customers, adding GPT-Live voice integration, prompt caching and structured JSON input. TechCrunch reports responses of roughly 100 to 180 milliseconds, yet public indexes still rank TypeSafe's Jev ahead.

By Rowan Ashby Published 22.10.26 23:15 UTC 6 min read

OpenAI / Wikimedia Commons / Public domain · Source

Key takeaways

OpenAI made its Decisions API generally available on Thursday, October 22, 2026, running on GPT-6 Luna through a dedicated /v1/decisions endpoint, according to the OpenAI developer changelog.
The Decisions API is a "decision model" product: instead of writing prose, it returns a probability, one option from a fixed list, or a score.
Since the September 29 developer-day preview and the October 6 public beta, OpenAI has added native voice integration with GPT-Live, prompt caching, structured JSON input and a 400-question limit per request.
TechCrunch reports response times of roughly 100 to 180 milliseconds, down from one to three seconds; Speak says the gap between actions fell from 1.4 seconds to 0.13 seconds.
On public indexes, OpenAI still trails TypeSafe's Jev, Liquid's D1 Ultra and Base Lab's Elephant-1, according to rankings cited by TechCrunch.

What is the OpenAI Decisions API?

The OpenAI Decisions API became generally available on Thursday, October 22, running on GPT-6 Luna through a dedicated /v1/decisions endpoint. Instead of writing prose, the decision model returns a probability, one option from a fixed list, or a score.

According to OpenAI's beta documentation, summarised by Let's Data Science and Unite.AI, there are three question types:

Predicate: returns a probability between 0 and 1 that a condition is true.
Choice: picks one value from options the developer declares, with a probability for each option and a separate confidence score.
Score: rates the input against ordered levels.

Netlify, which added the API to its AI Gateway, gives a simple example: a returns-routing function classifies a jacket with a broken zipper as "defective" with 0.95 probability and 0.93 confidence. Because the answer is always one of the declared options, the application can act on it directly, for example by sending a prepaid return label, and pass low-confidence cases to a human support agent.

This differs from a chatbot call. A tool-calling AI agent usually has to generate text and then parse it. A decision model skips the free-form text, which is why it can be much faster and cheaper per call.

What changed between the preview and general availability?

OpenAI first showed the product at its developer day on September 29, according to TechCrunch. It opened a public beta to all developers on October 6, and the changelog lists general availability on October 22. The main changes are below.

Item	Public beta (Oct. 6)	General availability (Oct. 22)	Source
Model	gpt-6-luna only	gpt-6-luna only	OpenAI changelog
Questions per request	Up to 200	Up to 400	OpenAI changelog
Inputs	Text and inline images	Adds structured data via `input_json` and images inside per-question instructions	OpenAI changelog
Prompt caching	Not available	Shared-input caching with prompt-cache discounts	OpenAI changelog
Voice	None	Native integration with GPT-Live (added Oct. 15)	OpenAI changelog
Typical response time	1 to 3 seconds (earlier approach)	About 100 to 180 milliseconds	TechCrunch
Price	$0.10 per 1M input tokens; output not billed	$0.10 per 1M input tokens, $0.01 cached; output not billed	Let's Data Science; Superpower Daily

TechCrunch describes the public release as adding voice and visual modalities. The changelog is more precise: image input already existed in beta, and voice works through GPT-Live, OpenAI's real-time voice model, which can call Decisions directly or trigger decisions as transcript text arrives. The changelog does not say that the Decisions endpoint accepts raw audio itself.

OpenAI also worked on location. TechCrunch reports that the company directed inference resources to specific regions to cut latency worldwide. Coverage of the launch says regional processing is available in the US and EU, and that Zero Data Retention is available to eligible customers.

One notable detail: OpenAI did not train a separate model for this. According to TechCrunch, it reconfigured its existing Luna model to run faster and cheaper. OpenAI's own speed claim, "10x faster than the Responses API," compares the endpoint with Luna through OpenAI's regular API, not with rival decision models.

How does it rank against Jev, D1 Ultra and Elephant-1?

OpenAI is entering a market it did not create. TechCrunch reports that the project began after OpenAI staff looked at Jev, the decision model from TypeSafe. API product lead Nikunj Handa told TechCrunch that his colleagues' reaction was that they had "like a million use cases for this."

The competition has moved quickly. TypeSafe released Jev 1.14 in mid-October, adding a bounding-box output type and what it calls small improvements across tasks, with the biggest gain expected in multilingual support, according to the company's blog. Liquid released D1 Ultra the week before OpenAI's launch, according to TechCrunch.

On public benchmarks, OpenAI is not yet the leader. TechCrunch reports that:

On the AIM System One Index, D1 Ultra and Base Lab's Elephant-1 rank highest, the latest Jev ranks 7th and OpenAI's decision model 8th.
On the Artificial Analysis Decision Index, the latest Jev ranks 3rd and OpenAI's model 14th.
On JevMan, a Pac-Man-style test built by Opper AI, the latest Jev tops the ranking.

The Artificial Analysis index is a weighted average of four tests, each worth 25%: knowledge and hallucination, multiple-choice reasoning, tool calling, and safety classification, according to its methodology page. It also tracks cost, time per task and calibration, meaning whether a model's stated confidence matches how often it is right.

Usage tells a similar story. OpenRouter's decision-model leaderboard, with data through October 21, shows TypeSafe far ahead by request volume:

Rank	Model	Developer	Requests
1	Jev 1.13	TypeSafe	948M
2	Jev 1.14	TypeSafe	316M
3	Decider V1.2	Perplexity	43.5M
8	GPT-6 Luna Decisions	OpenAI	21.8M
10	d1	Liquid	19.4M

Source: OpenRouter decision rankings, data through October 21, 2026. Figures cover traffic routed through OpenRouter only, not customers who call OpenAI or TypeSafe directly.

What are early customers reporting?

OpenAI named Box, Datadog, Ramp and Speak among its trial customers, according to TechCrunch. The results they shared are self-reported and specific to their own workflows:

Speak: CTO Andrew Hsu told TechCrunch that the time between actions in the language-learning app fell from 1.4 seconds to 0.13 seconds.
Box: route selection was about five times faster across 30 insurance-certificate cases, with the same accuracy as the Responses API, according to launch coverage citing OpenAI.
Ramp: AI cost attribution ran four times faster and 30% cheaper than standard GPT-6 Luna calls without losing accuracy, according to the same coverage.

OpenAI also uses the API inside its own Dots voice assistant to make it more responsive, TechCrunch reports. Handa told the outlet that "the big unlock with Decisions API has been the speed."

What it means

General availability matters for businesses that need stable terms before they put a model in production. OpenAI now offers a decision endpoint with input-only pricing, regional processing and voice support, sold by the same vendor many companies already use for prompt-based applications. That convenience is its main advantage.

The benchmarks suggest quality is not yet its advantage. Specialist models from TypeSafe, Liquid and Base Lab rank higher on the indexes TechCrunch cites, and TypeSafe dominates OpenRouter traffic. Developers comparing options may want to test accuracy and calibration on their own labelled data, as OpenAI itself recommends when choosing thresholds.

The launch also continues a wider push by large labs into specialised, cheaper endpoints, alongside moves such as OpenAI models arriving on Google Cloud and price competition seen in Gemini 4 Argon's pricing. What is not yet known is how OpenAI's model will score once indexes retest the general-availability version, and whether OpenAI will train a dedicated decision model rather than adapting Luna. OpenAI has floated 60-millisecond responses as a future goal, according to launch coverage, but that is not part of this release.

Frequently asked questions

What is the OpenAI Decisions API?

It is an OpenAI endpoint (/v1/decisions) that answers questions about an input with a typed result: a probability that a condition is true, one option from a declared list, or a score. It runs on GPT-6 Luna and does not generate free-form text.

How much does the OpenAI Decisions API cost?

According to OpenAI's beta documentation and launch coverage, it costs $0.10 per 1 million input tokens, or $0.01 for cached input. Output tokens are not billed, though regional processing premiums and long-context multipliers can apply.

How fast is the Decisions API?

TechCrunch reports typical responses of about 100 to 180 milliseconds, down from one to three seconds. OpenAI says it is about 10 times faster than using the same model through the Responses API.

Is OpenAI's Decisions API better than Jev?

Not on current public indexes. TechCrunch reports that Jev ranks 3rd and OpenAI's model 14th on the Artificial Analysis Decision Index, and Jev models lead OpenRouter usage by a wide margin.