The situation.
A message, document, or relevant facts. TypeSafe calls this the state. Think “the information on the desk.”
01 / The big idea
What goes into a decision?
Every request has two ingredients: information to examine, and questions about that information. It helps to keep them separate.
The question.
One specific judgment. You supply the categories, the scale, or the yes/no question.
The decision.
A usable value and probabilities. Your software decides whether to act, ask for help, or do nothing.[1][17]
THE FORM ANALOGY
Jev fills in a carefully designed form. The surrounding software decides what to do with that form.
02 / A visual walkthrough
How questions can run in parallel
Jev can evaluate independent questions about the same information together. Each question stands on its own. A later decision that needs an earlier answer still requires another step.[8]
Explore the workflow in 3D
01 · Shared context. The message and relevant policy go in together. Every question sees this same information.
Drag or use arrow keys to rotate · A conceptual workflow, not the model’s internal architecture.
Select a step to explore it.
03 / The decision playground
Choice, Score and Noul
Switch the question type, then change the situation. Watch the shape of the answer change.
“You billed me for two memberships. I only signed up for one.”
The question
Which team should handle this?
A Choice selects from categories you supply. The categories do not need to have an order.
The selected category
Billing
The bars share 100% between the available options. “Other” gives unusual requests somewhere to go.
Teaching simulation · Values are invented, not measured Jev responses.Read the specification Choice ↗
Probability asks: “How likely is this answer?” For Choice, all option probabilities add up to 100%. A Score also returns a distribution across the levels you described.[3][4]
Confidence summarizes the spread. Choice and Score also return a 0–1 confidence statistic derived from that distribution. It is not simply the winning probability or a guarantee of accuracy. Noul has no separate confidence field.[6]
04 / Understanding uncertainty
How to interpret
a probability
Imagine 100 cases, each given an 80% chance of “yes.” In a well-calibrated model, roughly 80 should actually be yes. The other 20 should be no.
That is calibration: matching probabilities to how often things happen across many cases. One prediction cannot prove it.[7]
TypeSafe calls its training approach RLCD—reinforcement learning for calibrated decisions. The aim is useful uncertainty, not just a confident-sounding answer.[7]
About 80 yesAbout 20 no
Idealized illustration of 100 cases, not a measurement of Jev’s calibration.
You decide when the software acts.
Try routing 12 example messages. Here, “yes” means the customer asks for a refund. Raising the threshold sends more uncertain cases to a person.
Auto yes at 85% or above. Auto no at 15% or below. Review everything in between.
0Automatic
0Human review
0Auto mistakes
Invented probabilities and known labels. These counts illustrate a trade-off; they do not report Jev’s accuracy.
Select a message to see its known label. An orange underline marks an automatic mistake.
The threshold here applies to the Noul probability. Real thresholds need testing against the cost of mistakes in your own task.[5]
05 / Where it fits
Where Jev fits in a larger workflow
Jev supplies a decision inside a workflow. Software still owns permissions, calculations, and the action itself.[15]
Get the message to the right team.
Identify the issue and whether the customer asks for a refund. Route the message; let a person or a separate system handle the response.
Customer message
Jev selects a team
Software routes it
A routing decision does not itself issue a refund.
Examples adapted from TypeSafe’s use-case map; these are workflow ideas, not built-in end-user apps.[18]
06 / Numbers, with context
Cost, speed and the evidence
Pricing is a published specification. Speedups depend on the test. Here is how to interpret both.
Published direct API input price · USD
$0.042
Tokens are the small pieces of text a model reads. The input includes context and questions. More text costs more; “one decision” does not have a universal price.
Example calculation
Estimated model input cost / month
100,000 × 1,000 ÷ 1,000,000 × $0.042
A price illustration. Excludes retries, other services, taxes, and provider-specific charges. Token limits still apply.
How to read the benchmark claims
193.6×
faster, in TypeSafe’s workflow evaluations
444.6×
cheaper, in those same evaluations
These are company-reported comparisons. TypeSafe says they may be near the high end of real-world gains. Its team authored the workflows, used other models’ probabilities as a reference, and required comparison models to produce probabilities too. These choices matter.[11]
This guide has not independently reproduced those results. A short routing task, a long document, and a request from another region can have very different timings.
The current technical details, in plain English
| Model | Jev 1.13.0. The jev-latest alias currently points to it. |
|---|---|
| What it reads | Text, including text organized as structured records. No direct image, audio, or video input. |
| Room for context | 64k tokens for the whole request; the shared state plus the longest single question must also fit within 32k. |
| Choice / Score | Up to 255 choices; 2–10 described Score levels.[3][4] |
| Published rate limits | 1,200 requests per minute; 250,000 tokens per second. TypeSafe says these can change while access expands. |
| Language | English is currently strongest; test other languages on your own material. |
| Customization | Supply examples and criteria. TypeSafe says there is no customer-specific fine-tuning. |
Direct API limits and aliases from the model documentation.[9] Provider interfaces may differ: OpenRouter lists a 32k context window.[16]
07 / Limits and common misconceptions
Why “type-safe” does not mean “correct”
If the allowed answers are Billing and Delivery, Jev stays inside that menu. It can still send a billing problem to Delivery. Type safety is a guarantee about the form of the answer, not the truth of the decision.[15]
It can be misled.
Ambiguous wording, missing facts, and malicious instructions hidden inside the supplied text can change the answer. A decision model is not an infallible security gate.
Keep exact math in code.
TypeSafe documents weaknesses in counting, numerical precision, and date comparisons. Let ordinary software calculate those.
More context can hurt.
Irrelevant material can reduce accuracy. Independent questions do not remove the problem of an overloaded input.
It will not write the explanation.
Jev does not generate a reply or a reasoning essay. Use a generative model when the job needs prose or extended reasoning.
TypeSafe explicitly documents these limitations for Jev 1.13.[10]
TAKEAWAY
Give it a narrow job, test it on real examples, and keep a route to human review.
08 / Check your understanding
Apply what
you’ve learned
Four quick situations. Choose an answer to see why it fits.
Your score
A little more background
Background and
practical questions
Expand a question for more detail, including who built Jev and how to try the real service.
Who is behind Jev?
TypeSafe AI’s founders are Diogo Almeida (CEO), Sasha Sheng (COO), and Erik Gafni (CTO). Their backgrounds include AI research and production systems at OpenAI, Google Brain, and Meta/FAIR. The company announced Jev’s early access on 15 September 2026.[12][11]
Why “System One”? And why “Jev”?
System One borrows from the idea of fast, intuitive thought popularized by Daniel Kahneman. It describes the intended task, not a claim that the model has a human mind.[2] Jev is named after economist William Stanley Jevons: TypeSafe’s bet is that cheaper intelligence will unlock more uses for it.[11]
How is this different under the hood?
TypeSafe describes a new architecture, a parallel sampler, and RLCD training. The useful consequence: many independent judgments can come back together without writing an answer token by token.[11] Our 3D diagram explains that workflow; it is not a reconstruction of the network.
Can it browse the web or remember my business?
You supply the relevant information with the request. Your application must fetch documents, keep records, and decide which context to send. It can include your policies and examples in that context.[17]
What happens to the data I send?
The real hosted service receives the context you submit. TypeSafe says it does not train on customer requests or responses, and offers zero data retention for enterprise customers. “Not used for training” is not the same as “not retained”; check the applicable data agreement.[14] The simulations on this page send nothing to Jev.
Can I try the real thing?
Start with TypeSafe’s console and quick start. Developers can use the HTTP API or Python and JavaScript libraries.[13] Jev also appears in Vercel AI Gateway and OpenRouter; those interfaces may use different names or limits.[15][16]