OpenAI Decisions API vs Jev: Three Answers Apart in 400

OpenAI's Decisions API, which runs the model GPT-6 Luna, opened in public beta on 6 October 2026. On 8 October we added it to DecideBench, our test of 400 decisions like whether a refund falls inside a 30-day window. TypeSafe's model Jev got 392 right and Luna 389. Sending worked examples, one solved example per answer with each question, Jev cost $32.26 per million decisions and Luna $52.82.

If you already build on OpenAI and want one bill, one contract and US or EU data residency, use Luna. It is three answers behind on 400, which is too close to call.

If you send worked examples and want the lower bill, use Jev. It is cheaper per decision and a little more accurate.

If you need to decide on images, use Luna. Jev reads text only.

If speed is the deciding factor, test both from your own servers. Our latency figures cannot settle it.

• Who makes them: OpenAI makes Luna, its only model on the Decisions API. TypeSafe makes Jev, which launched on 15 September 2026.
• What they do: both read a piece of text, a question and a short list of answers, and return one answer with probabilities, not prose.
• Accuracy: Jev 98.0%, Luna 97.2%. Each got at least 47 of 50 right in each of the eight kinds of task we test.
• Price: input only. Jev charges $0.042 per million tokens, Luna $0.10. A token is about three quarters of a word.
• Status: Luna is a beta, with general availability expected "in the coming weeks". Jev's own API has a waitlist, and Cloudflare also sells it.

New to the category? What is the OpenAI Decisions API explains how Luna works, and what is Jev covers the model it competes with.

What changed this week

OpenAI's beta went public on 6 October 2026, per Unite.AI. It had been a limited preview since DevDay 2026. The Decisions guide prices Luna at $0.10 per million input tokens, with no charge for output, cache reads or cache writes.

Two days later we ran it on DecideBench, with and without worked examples, and published both results in the changelog.

DecideBench is our own benchmark, and we publish it so anyone can check it. It has 400 multiple-choice decisions in 200 pairs. The two halves of a pair differ by one small edit that changes the right answer, such as a return request one day past the window. A model that matches on topic words gets one half wrong. How to test a classifier explains why we built it that way.

Which is more accurate, OpenAI's Decisions API or Jev?

Jev is more accurate by three decisions in 400, a gap too small to call either model better. Jev got 98.0% and Luna 97.2%. Counting only pairs where both halves were right, Jev got 192 of 200 and Luna 189.

The worked examples did little for either model. Every model that can take them sees one solved example per answer before each decision. Without them, Luna scored 97.0% and Jev 98.2%, a gap of five decisions. Neither model needs examples to do this job well.

The eight task families tell the same story. Each family has 50 decisions, so a two-point gap is one decision.

Jev is one decision ahead in four families and Luna in one, checking claims. The other three are level. Both models do worst on returns, where the answer turns on a date or a policy clause.

Three larger general chat models score higher still, between 99.2% and 99.8%, at six times Jev's cost or more. Decision model vs LLM works through when those last points are worth paying for.

Which costs less, OpenAI's Decisions API or Jev?

Jev costs less when you send worked examples: $32.26 per million decisions against Luna's $52.82. Without examples, Luna's bill falls to about $25 per million, by the benchmark's estimate.

Each model charges for input only. Jev's price per token is less than half of Luna's. Jev's API bills more tokens for the same decision, about 770 against Luna's 530 by our arithmetic from the bills, but the lower rate wins.

The no-examples row is where Luna catches up. Luna's input drops from about 530 tokens to 245 per decision. DecideBench did not record Jev's bill for its no-examples run, so we cannot say which is cheaper there. Without examples Jev also reads fewer tokens, so its bill would fall too, by an amount we did not record.

At ten million decisions a month, the gap is about $200: $323 on Jev, $528 on Luna with examples. The gap is real money for a high-volume screening job and a rounding error for most teams. OpenAI's guide adds that regional processing premiums and long-context multipliers can raise Luna's rate.

Our own bills for Jev on real work, including a $26.63 run over 143,891 items, are in Jev pricing.

Which is faster, OpenAI's Decisions API or Jev?

Luna answered faster in our runs, but the two numbers were not measured the same way, so we do not count the gap as a result. Luna's median was 126 ms and Jev's 639 ms.

• Luna was timed from a cloud datacenter.
• Jev was timed from our client machine, network included, through AI Space, our own gateway. That extra hop is in its time.

OpenAI and TypeSafe each make speed claims of their own. OpenAI says the Decisions API is "about 10x faster than the Responses API", and gives no figures. TypeSafe says Jev's end-to-end response time is 70 to 500 ms, measured from a laptop in the western US. Both are the vendors' own claims.

The fix is to time both from the place your software runs.

How do the two APIs differ?

Jev and Luna both answer a fixed-format question about a piece of text and return probabilities. Luna also takes images, and only Jev's maker says its probabilities are calibrated, meaning a 90% answer is right about 90% of the time. The request shapes differ enough that switching means rewriting the call. What is the OpenAI Decisions API walks through Luna's request in full.

Calibration is the difference that matters most in production, because it lets you act on confident answers and send the rest to a person. TypeSafe's launch post says every Jev answer comes with "calibrated probabilities and confidence scores". OpenAI's guide makes no such claim. It tells you to set thresholds from labelled examples from your own application, weighing what a false yes costs against a false no.

Setting thresholds from your own labels is sound advice for both models. A vendor's calibration claim is a starting point, and your own labelled sample is what tells you where to set the line. LLM cascade shows how we chose thresholds for Jev.

On data handling, OpenAI lists Zero Data Retention and HIPAA use for eligible customers, with processing in the US or Europe. Cloudflare lists Jev with zero data retention.

Which should you pick?

Pick by where your stack already lives, not by the benchmark gap. Three decisions in 400 will not decide this for you.

• Pick Luna if you already have an OpenAI contract, need images, or need a named data region. It is a beta, so plan for the request format to change before general availability.
• Pick Jev if you want the lowest bill we measured with worked examples, a vendor calibration claim, or more than one place to buy it.
• Run both if you can. Both bill only for input, so a few hundred of your own labelled items cost cents on either. That sample tells you more than our 400.

If neither fits, the best Jev alternatives covers open models you can run yourself, and Clef vs Jev covers Cloudflare's open pair.

What this comparison cannot tell you

• The accuracy gap is noise. Three decisions in 400 could flip on a re-run or a different set of tasks.
• The latency numbers are not comparable. Luna was timed from a datacenter, Jev from our client through our gateway.
• We did not test images. Luna's image input is untested by us, so we make no claim about it.
• Luna is a beta. OpenAI expects general availability in the coming weeks, and price or behaviour may change by then.
• A language model wrote the tasks. Claude wrote all 400, and a careful reader could dispute a few labels. Both models are measured against the same labels.
• No Jev bill without examples. The table marks it as not measured rather than guessing.

FAQ

Is OpenAI's Decisions API better than Jev?

Not on our test. Jev got 98.0% of DecideBench's 400 decisions right and Luna 97.2%, three decisions apart, which is too close to call. Jev cost less with worked examples, $32.26 against $52.82 per million. Luna is the better fit if you need images or already build on OpenAI.

How much does the OpenAI Decisions API cost?

Luna costs $0.10 per million input tokens, with no charge for output, cache reads or cache writes, per OpenAI's guide. On DecideBench that came to $52.82 per million decisions with worked examples. Without examples, each decision used about 245 tokens instead of 530, which is about $25 per million.

Is GPT-6 Luna faster than Jev?

Its median was 126 ms on our test against Jev's 639 ms. We timed Luna from a cloud datacenter and Jev from our own client through a gateway, which is not a fair race. Given TypeSafe's own claim of 70 to 500 ms for Jev, time both from where your software runs.

Does Jev accept images like the Decisions API?

No. Jev reads text or structured data only, and TypeSafe's launch post says it does not take images yet. Luna accepts images as inline base64 data, alongside text, in one message. It does not accept hosted image URLs or uploaded file IDs.

Are the Decisions API's probabilities calibrated?

OpenAI does not claim they are. Its guide calls the yes-or-no probability "an estimate" and recommends setting thresholds from your own labelled examples. TypeSafe does claim calibration for Jev. Either way, check on a labelled sample of your own before you let a threshold run unattended.

Sources

• Unite.AI, OpenAI Releases Decisions API in Public Beta, Powered by GPT-6 Luna, 6 October 2026
• OpenAI, Decisions guide, accessed 14 October 2026
• TypeSafe, Introducing System One Models and Jev, 15 September 2026
• Cloudflare, Jev model page, accessed 14 October 2026
• XY Space, DecideBench dataset and results, v1.1, accessed 14 October 2026
• XY Space, DecideBench leaderboard, accessed 14 October 2026
• XY Space, DecideBench changelog, OpenAI Decisions added 8 October 2026