Jev vs OpenAI Decisions API: Speed, Cost and Calibration Compared

Introduction

In September 2026, two products arrived within two weeks of each other, built on the same idea: AI that returns decisions instead of text.

Jev from TypeSafe AI, launched September 15
OpenAI Decisions API, announced September 29 at DevDay

They look similar from the outside. Underneath, they are quite different.

This guide compares them on what matters, and is explicit about what is measured and what is only claimed.

The Short Version

Jev is a purpose-built decision model: available now, documented, cheap, with calibrated probabilities. Text only.
OpenAI Decisions API is a constrained version of GPT-6 Luna: multimodal and backed by OpenAI's ecosystem, but in limited preview with no published pricing or schema.

Today you can build on one of them. The other you can only plan for.

Comparison Table

	Jev	OpenAI Decisions API
Launched	Sept 15, 2026	Announced Sept 29, 2026
Availability	Generally available, SDKs published	Limited preview
Latency	P50 ~0.21s, P95 ~0.34s (measured on OpenRouter, Oct 1)	~150 ms (OpenAI claim)
Input price	$0.042 per million tokens	Not published
Output price	$0	Not published (Luna: $0.50 per million)
Inputs	Text only	Text and images
Context window	32K (OpenRouter) / 64K (TypeSafe docs)	Not published
Probabilities	Calibrated, per option	Undocumented
Architecture	Purpose-built "System One" model	Constrained GPT-6 Luna
Can generate prose	No	In principle yes, but constrained

Speed

OpenAI claims about 150 ms per decision, roughly 10x faster than a normal Luna call.

Jev's numbers come from production traffic on OpenRouter: about 210 ms at the median, 340 ms at P95 and 750 ms at P99.

Both sit in the same range. The difference is evidence: Jev's figures are measured on real traffic, while OpenAI's is a single number from a keynote slide with unstated conditions.

For most marketing workloads, such as moderating comments or routing DMs, either is fast enough. Tail latency (P95, P99) matters more than the headline, and only Jev has published that.

Cost

Jev: $0.042 per million input tokens, with free output tokens.

OpenAI Decisions API: no Decisions-specific price has been published. The model underneath, GPT-6 Luna, lists $0.10 per million input and $0.50 per million output tokens.

If Decisions is billed like Luna, Jev would be around 58% cheaper on input, and output cost would favor Jev even more. But that is a guess, not a fact.

Why output cost matters: decision responses are short but structured, and output pricing can dominate when you make millions of calls.

Calibration: The Real Difference

This is where the two diverge most.

Jev is trained with RLCD (Reinforcement Learning for Calibrated Decisions). Its probabilities are optimized against real outcomes, so an answer scored 0.80 should be right about 80% of the time.

OpenAI Decisions API is described as asking a general-purpose model to pick from bounded answers. Whether it returns probabilities at all is unconfirmed, and if it does, self-reported model confidence is not the same as calibrated confidence.

Why it matters: calibrated probabilities let you build rules like "auto-approve above 0.95, send to a human below 0.7". Without trustworthy confidence, you cannot safely automate those thresholds.

For marketing, this is the difference between automation you can trust and automation you have to babysit.

Inputs

OpenAI wins here. Decisions accepts text and images, which suits agents that need to look at a screenshot or a creative before deciding.

Jev is text-only. For visual content you would need to describe or caption the image first.

If your workflow decides on images, such as checking whether a generated creative fits brand guidelines, this is a real limitation for Jev today.

Architecture

OpenAI constrains an existing generative model: it bounds the answer space before inference. That brings strong general knowledge, but the model underneath can still generate text.

Jev was built as a specialist. It was never trained to write prose, so it cannot drift off-schema.

Jev's documented primitives:

Choice: pick from options, with probabilities
Score: position on a rubric
Noul: probability that a condition holds

You can mix them in one call.

OpenAI has not documented specific primitives.

Availability and Developer Experience

Jev: generally available, Python and JavaScript SDKs, schemas published, usable through OpenRouter
OpenAI: limited preview, no public endpoint, SDK method or schema, and standard API keys reportedly return a "not enabled" error

OpenAI said a broader release was coming within days of the announcement. Until then, you cannot test it.

Which Should You Choose?

Choose Jev if:

You need to ship now
You make high-volume text decisions and cost matters
You want calibrated probabilities for approval thresholds

Wait for OpenAI Decisions if:

Your decisions depend on images
You are already deep in the OpenAI ecosystem
You want one vendor for both generation and decisions

For most teams, the best answer is both. Decision models are becoming a standard layer, and the sensible approach is to keep your system model-agnostic so you can swap or combine them as the evidence comes in.

What This Means for AI Marketing Teams

Whichever you pick, the pattern is the same:

Generative models write, design and plan
Decision models classify, route and gate quickly and cheaply
Humans handle the cases the system is unsure about

This is how autonomous marketing becomes both faster and safer, and it fits the multi-agent approach to building AI teams.

A Note on Evidence

As of early October 2026:

Jev's latency comes from third-party production telemetry; its accuracy and speed-versus-LLM claims otherwise come from TypeSafe
OpenAI's figures come from its own DevDay presentation
There is no independent head-to-head benchmark yet

Re-check the numbers when OpenAI publishes pricing and documentation.

Frequently Asked Questions

Which is faster, Jev or OpenAI Decisions API?

Both are in the 150 to 350 ms range. OpenAI claims about 150 ms; Jev measures about 210 ms median and 340 ms at P95 on OpenRouter. Only Jev's numbers come from production traffic.

Which is cheaper?

Jev costs $0.042 per million input tokens with free output. OpenAI has not published Decisions pricing; the underlying GPT-6 Luna costs $0.10 input and $0.50 output per million tokens.

What is the difference in probabilities?

Jev is trained to produce calibrated probabilities. OpenAI's Decisions API has not documented its confidence output, and self-reported model confidence is not necessarily calibrated.

Can either handle images?

OpenAI Decisions supports text and images. Jev is text-only.

Can I use them today?

Jev is generally available. OpenAI Decisions is in limited preview.

Final Thoughts

Jev and OpenAI Decisions are two answers to the same question: how do AI agents make fast, reliable choices?

Jev has the edge today on availability, cost transparency and calibration. OpenAI has the edge on image input and ecosystem reach, once it ships.

The smart move is not to pick a winner. It is to build systems that can use whichever decision layer proves best.