guide

Did OpenAI just kill Jev AI? Decisions API vs Jev

Jev publishes probabilities and a price. OpenAI announces image support. Compare the documented differences and decide what to test on your own workload.

EiliyaSeptember 30, 20267 min read

Jev AI publishes a price: $0.042 per million input tokens, with output free. OpenAI's Decisions API targets the same classification, routing and agent-action jobs, but its separate price is still unconfirmed.

At DevDay on September 29, Simon Willison described OpenAI's announcement as sounding like a "response to Jev." That's his interpretation of OpenAI's motives. If you're comparing the products on September 30, 2026, Jev already has a published interface and probabilities to inspect; the useful questions are what you can access, which outputs and inputs you need, and how each performs on your workload. Read Willison's live blog.

What Jev AI returns to your code

Jev is TypeSafe AI's first public System One model. The company's launch post, dated September 15, describes a system built for structured decisions rather than generated prose. You supply application state and questions. Jev returns judgments your code can use.

TypeSafe's question documentation defines its output types:

Choice selects an option and returns probabilities plus confidence.
Score evaluates a defined scale and returns a score with its probability distribution and confidence.
Noul returns the probability that a yes-or-no statement is true. It has no separate confidence field.

Questions about the same state can run in parallel. Your support app could ask which queue owns a ticket and how urgent it is in the same request, then combine those answers into its handling policy.

The company's Business Wire announcement reports a $40 million seed round led by DCVC and names founders Diogo Almeida, Erik Gafni and Sasha Sheng.

Decisions API vs Jev: the documented differences

OpenAI's DevDay recap describes a Luna-based service with finite answers, text or image context, and limited preview access. Broad release is planned soon. The announcement does not supply a separate price or detailed public response schema.

Read this as a comparison of what each company documents. A blank in the docs is a question to ask, not an invitation to guess.

Point	OpenAI Decisions API	Jev
Underlying model	Luna	TypeSafe's System One model
Decision interface	User-defined questions with finite answers	Documented Choice, Score and Noul questions
Inputs	Text or images, per the recap	Text only, including structured text
Probabilities	Not specified in the recap	Documented distributions and yes/no probabilities
Multiple questions	Detailed execution behavior not confirmed	Parallel questions over shared state
Latency evidence	Keynote claim of subsecond responses	TypeSafe reports 70 to 500 ms end to end
Price	Separate tariff not confirmed	$0.042 per million input tokens; output free
Access	Limited preview; broader release planned	Early access and published Vercel gateway integration
Public integration detail	Dedicated reference not confirmed in our research	HTTP API and SDK documentation

Jev's input and price rows come from its current models documentation. TypeSafe supplies the timing claim in its launch materials. The table gives you no head-to-head accuracy result.

Where Jev still has an edge

Jev's advantage today is a decision contract you can inspect. You can see which outputs exist and how uncertainty reaches your code. OpenAI's announcement doesn't yet document an equivalent probability or confidence interface.

For an ambiguous request, a chosen label alone may not tell your router whether to act at all. Jev's confidence documentation explains how to examine the distribution and choose an escalation policy. Test those thresholds on the task you'll actually run.

You also have a documented path into an existing gateway stack. Vercel's integration announcement describes access through its AI SDK, HTTP API or an existing TypeSafe client. You can plan around that integration detail; OpenAI's preview still lacks a confirmed public request example.

Parallel questions may help when your app needs several independent judgments about the same input. Try your actual questions. If a later decision needs information retrieved because of an earlier result, parallel evaluation can't skip that dependency.

These implementation advantages could narrow as OpenAI publishes more detail. For now, they give you reasons to evaluate Jev rather than assume OpenAI has displaced it.

OpenAI's image input is worth testing

Image context is the clearest announced difference. Jev's current documentation requires you to convert non-text inputs into text or structured fields. OpenAI explicitly supports images as context for Decisions API.

If your router depends on a screenshot or a scanned document, direct visual input could simplify the pipeline. Test that benefit: OpenAI hasn't supplied a reproducible accuracy comparison for that task in the sources reviewed.

If your team already buys and operates OpenAI APIs, evaluating a Luna-based service within that supplier relationship may be easier administratively. That depends on your agreement, access and operational requirements. Decisions-specific enterprise controls and service commitments remain unconfirmed.

OpenAI's planned broader rollout could make access easier. Check your account before making the service a dependency in your release schedule; a rollout plan doesn't give you access.

Compare the bill, then question the speed claims

Jev's published input price is clear. Decisions API's price is not. The OpenAI pricing page lists regular GPT-6 Luna standard short-context usage at $0.10 per million input tokens and $0.50 per million output tokens. Those rates are not a confirmed Decisions API tariff.

Using Luna's ordinary rates as Decisions pricing would put an unsupported number in your comparison. Without a Decisions tariff, neither service has a settled cost advantage.

Compare the bill for completing the same decisions at comparable quality. Include low-confidence fallbacks, failed requests and any input preprocessing. A cheap classification that sends work to the wrong handler has costs outside the model invoice.

TypeSafe's speed multiples need context too. Its launch materials say timings were generally measured from West Coast laptops. The company acknowledges that short input favored its demonstration and that its workflow comparisons likely represent the higher end of real-world gains.

Our take: compare Jev with a short, strict classification call as well as a reasoning-heavy frontier model. The latter is a valid baseline if your task needs it. It is a poor default for judging every queue assignment.

The 13% adoption figure and a Pokémon demo

The Hacker News launch discussion mixes interest with skepticism. In an individual comment, user jacobgold questioned the speed comparison and separated avoiding invalid types from choosing the correct valid answer. Keep that objection in mind when testing either product.

Vercel reported that nearly 13% of its paid teams used Jev within its first day on AI Gateway. That's evidence of use beyond launch copy, measured by Vercel itself. It gives you neither a named customer list nor evidence of long-term retention.

Read that figure alongside the launch promotion. Vercel's Academy course page says free promotional usage ended September 25, 2026. Trying a model for free doesn't prove teams will keep paying afterward.

The public jev-pokemon repository shows an actual implementation: its game wrapper reads memory, lists legal options with facts and lets Jev select one. You can inspect the bounded decision loop. The demo doesn't establish enterprise routing performance or an ability to understand arbitrary visual interfaces.

A named production customer roster wasn't confirmed in the primary materials reviewed. Treat gateway adoption and public projects as evidence of experimentation, without turning them into customer endorsements.

Test one decision before choosing a provider

Choose a decision you can label and replay. Give both services the same relevant context and answer definitions, then compare correct outcomes, delay and cost. Keep a strict small-model baseline in the test too.

If you need probabilities for text routing today, Jev has the more inspectable public interface. For visual decisions, OpenAI's announced image support is a reason to test its preview when you get access. Your existing router may still be the best option until either challenger proves a practical benefit.

Software needs plenty of judgments that don't require a paragraph. Both products address that problem, and the evidence doesn't establish a winner. OpenAI has challenged Jev's category; a claim that it has killed the competitor runs well ahead of the evidence.

FAQ

Is Jev AI the same as OpenAI Decisions API?

No. Jev is TypeSafe AI's model. Decisions API is OpenAI's Luna-based preview. They overlap in classification and routing but have different documented interfaces and input support.

Is Jev cheaper than Decisions API?

There's no confirmed comparison yet. Jev publishes input-only pricing, while a separate Decisions API price remains unconfirmed. Don't substitute ordinary Luna rates for it.

Can Jev make a wrong decision?

Yes. A typed answer can still select the wrong valid option. TypeSafe documents weaknesses in numerical tasks, literal interpretation and adversarial content.

Should developers switch from Jev to OpenAI?

Switch when you have access and tests on your own task show a benefit. OpenAI's image input may matter; Jev's probabilities and documented integrations may matter more for a text router.