OpenAI's Decisions API, side by side with Jev

At its DevDay keynote on September 29, OpenAI previewed a Decisions API. You give it questions and a fixed set of possible answers, and it returns one of them. It is built to hand your code a usable answer quickly.

It belongs to the same family as TypeSafe’s Jev, which we covered yesterday. OpenAI shipped its own decision model about two weeks after Jev appeared, so developers paid close attention. Every went as far as saying that, of the five DevDay announcements it singled out, this one could turn out to be the biggest.

There is not much to go on yet. OpenAI’s official description is a single paragraph in its DevDay recap, and there is no price. Below is what has been confirmed so far, the first head-to-head tests, how it compares with Jev, and what to check when it opens up.

At a glance

Item	Details
What it does	Picks an answer from questions and options you define
Model	Luna, the smallest and most affordable model in OpenAI’s lineup
Input	Context as text or images
Response time	About 150 ms (OpenAI, as reported by The New Stack)
Status	Limited preview, with a broad release “in the coming days”
Price	Not announced

What OpenAI has said

This is how OpenAI’s DevDay recap describes it:

Decisions API enables real-time decision-making by focusing Luna’s intelligence on a specific set of user-defined questions with finite pre-defined answers. Developers supply context using text or images, and get back answers they can use to classify content, route requests, or choose an agent’s next action.

Simon Willison, live-blogging the keynote, noted that it works by giving the Luna model “a predefined set of options to choose from” and responds “in a fraction of a second”. The New Stack quoted OpenAI as putting the response time at 150 milliseconds, against 1.6 seconds for the same decision made with a regular GPT-6 Luna call.

Much is still missing. As of the morning of September 30 (Korean time), OpenAI’s developer docs, API changelog and pricing page had no entry for the Decisions API. An OpenAI spokesperson told The New Stack that the company plans to share more at broad rollout, so this post shows no request format.

One naming note: OpenRouter has an alpha endpoint also called the Decisions API, which serves Jev. It is a different product.

Where it fits

OpenAI names three uses:

Classifying content: is this message spam, and what topic is it about?
Routing requests: which team, or which model, should handle this ticket? This is the router in front of your LLMs from our previous post.
Choosing an agent’s next action: which of the buttons on screen should it click? Every’s example of deciding whether an email needs a reply fits here too.

Image input is the most visible difference from Jev. Jev reads text only, so screens and photos first have to be turned into text by another model. With the Decisions API you could pass a screenshot as context and ask which button to press next. How well that works will only be clear once more people can try it.

The first head-to-head tests

Every published results from testing the preview side by side with Jev.

Test	Decisions API	Jev
Replay of computer tasks (text only), correct steps out of 78	76	73
Typical response time on that test	230 ms	500 ms
Accuracy classifying conversation threads	Effectively tied	Effectively tied
Median response time on that test	309 ms	161 ms

The first two rows come from Every’s Jack Cheng, the last two from Kieran Klaassen of Cora. Each test favored a different side. Every’s summary: the Decisions API did better than Jev on speed and accuracy in a few early tests but trailed on its broader evaluations.

The samples are small and the conditions were not published in detail. Treat the results as a reason to run your own comparison rather than as a verdict.

Side by side with Jev

Both return one of a set of predefined answers and are built for decisions only. With what is public today:

Item	OpenAI Decisions API	TypeSafe Jev	Source
Status	Limited preview, broad release soon	Early access (waitlist)	OpenAI recap, TypeSafe
Input	Text, images	Text only	OpenAI recap, TypeSafe docs
What comes back	The chosen answer with a confidence score	The chosen answer, a probability per option, confidence	The New Stack, TypeSafe docs
Price	Not announced	$0.042 per million input tokens, output free	The New Stack, TypeSafe docs
Response time	About 150 ms (OpenAI)	70 to 500 ms (TypeSafe)	The New Stack, TypeSafe
Option limit	Not announced	255 per Choice question	The New Stack, TypeSafe docs
Where it runs	Not announced	US West Coast	TypeSafe

The biggest differences are image input and a published price. Teams already on the OpenAI API may like not adding another vendor. Until the price and limits are out, though, nobody can work out what it costs to run.

What to keep in mind

Price is the deciding factor. Decision calls are small and frequent, so the price decides where they make sense. Every’s view is that if the API is cheap enough for thousands of little decisions, it could be one of DevDay’s biggest releases.
Option limits and tuning are unknown. The New Stack points out that it is unclear how many candidate answers one request can take and whether developers can tune it on their own data.
There is no evidence on calibration yet. The value of a decision model is a probability you can trust, and OpenAI has not published a calibration evaluation for this API. Score it on records where you know the right answer before you set thresholds.
Speed depends on the setup. OpenAI says 150 ms; Every’s two tests saw 230 ms and 309 ms. OpenAI has not said where it runs, so measure from where your users are. When we measured Jev from Korea, the network round trip alone added 0.14 to 0.17 seconds.
The launch timing differs by source. OpenAI’s recap says “in the coming days”; Every wrote “in a few weeks”.
Resistance to manipulated context is unknown. Research on Jev showed that a few plausible sentences can flip its decisions. Nobody has yet shown how the Decisions API holds up to the same trick.

And in endue

The principle from our previous post holds whichever decision model you choose: make the cheap, fast decision first and call a large model only when needed. In endue you can apply it by setting a light model as an agent’s default and switching to a stronger one only for the messages that need it. The settings are covered in Choosing a model.