~/wiki/entities

Jev (model)

[ entity ][ updated 2026-09-17 ][ confidence high ][ jev-1.13.0 ]#jev · model · pricing · limits · system-one

TL;DR Jev is TypeSafe's flagship and first System One model. Current version jev-1.13.0 (aliases jev-latest, jev-preview). Text-only input, typed decision output, $0.042 per million input tokens with output free, 64k total context (32k for state plus the longest question), 250,000 tokens/sec and 1,200 requests/min rate limits. It cannot generate text and by construction cannot return a value outside your schema — but it can return the wrong valid one.

Facts

Field	Value
Name	Jev — named after William Stanley Jevons ("a nod to Jevons Paradox")
Model class	System One model (System One Models)
Current model ID	`jev-1.13.0`
Aliases	`jev-latest` → `jev-1.13.0` (most recent stable release; SDK default); `jev-preview` → `jev-1.13.0` (most recent release, stable or not; currently identical — "There is no preview build available right now")
Price	$42 per Btok / $0.042 per Mtok, charged on input tokens only; output tokens are free ("too cheap to meter")
Rate limits	250,000 tokens per second; 1,200 requests per minute. Over either → `429 Too Many Requests`. Docs warn limits "can change without notice"
Context length	64k tokens per request (state + all questions); 32k tokens for `state` plus the single longest question
Modality	Text only — string, JSON object, or array of text values. No image, audio, or video input
Endpoint	`POST /v1/systemone` (all models share it); `GET /v1/models` lists the aliases
Latency (claimed)	70 ms – 500 ms end-to-end, versus "3 to 329 seconds" claimed for frontier LLMs
Availability	Early access, waitlisted, since 2026-09-15
Customization	None per account: not fine-tuned or LoRA-adapted with customer data; "the same weights serve every account"
Training	RLCD — Reinforcement Learning for Calibrated Decisions
Data handling	"Jev is not trained on customer requests or responses"; ZDR available for enterprise
Languages	Accepts natural-language text; English is "the primary training language and where accuracy is currently best"; other languages including CJK "handled but not equally well"

What it does

Jev evaluates typed questions against one state and returns structured results — no text generation, no parsing. Three primitives (Primitives: Choice, Score, Noul):

Question type	Goal	Returns
Choice	Choose an option from a list	`choice`, `probabilities`, `confidence`
Score	Score the state on a rubric	`score`, `probabilities`, `confidence`
Noul	Is this statement true?	`noul` (0–1)

All three types can be mixed in one call. "Every question is evaluated in parallel and in isolation against the same state in one go. Adding questions barely changes the response time… so adding more questions does not create context-rot" (raw/docs/introduction.md). Jev ingests the state once and evaluates every question against it in parallel, which is why the context budget is split the way it is.

The launch blog's framing: "Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out." Sampling is parallel rather than autoregressive: "Generates all outputs in a single query."

What it does not do

No string generation. "While Jev gives up string generation, it's optimized for structured outputs and can't hallucinate" (blog). Concretely: possible outputs and structure are defined in advance and "The model never makes type errors." The blog is explicit that the 0% type-error figure "is not empirical. Schema matching is guaranteed, thus we can confidently add 0% into the plots."
No arithmetic, counting, or date ordering. "Jev is not a calculator… jev-1.13 does not count reliably." Dates should be extracted as components and compared in code.
No indirection-heavy reasoning. jev-1.13 "may struggle with tasks that require additional levels of indirection. It can be quite literal in its understanding. It struggles with tasks that require numeric precision."
No lookups. Jev only knows the state you send.

The nine documented failure modes of jev-1.13 — literal reading, math and numbers, date and time comparison, indirection, large state full of irrelevant detail, adversarial content, contradictory instructions and criteria, common-sense structural invariants, and generation — are catalogued in Jev 1.13 jaggedness: known failure modes (jaggedness page "Last reviewed 2026-09-17").

Note the distinction the sources keep: schema safety is guaranteed; judgment is not. Jev can return a wrong but valid answer.

Version history

The docs do not publish a model changelog. What the sources establish:

Version	Evidence
`jev-1.12`	Used in cookbooks whose results are dated 2026-08-11 and 2026-08-12 (`TYPESAFE_MODEL = "jev-1.12"`, "$ per 1M tokens (input, output); TypeSafe jev-1.12 as of 2026-09")
`jev-1.13` / `jev-1.13.0`	Current release as of 2026-09-17; jaggedness page scoped to `jev-1.13`; both aliases resolve here

Aliases move when a new release ships, so "the answers behind it can change without a change on your side." The response's model field reports the versioned ID that answered — pin the version if you have tuned confidence thresholds. See Versions and timeline (models, SDKs, API, company).

System One Models — the model class
Models, aliases, pricing, rate limits, context — the full contract for prices, aliases, and limits
HTTP API: POST /v1/systemone and GET /v1/models — request and response shapes
Jev 1.13 jaggedness: known failure modes — failure modes in detail
Confidence vs probability — how to use the confidence channel
TypeSafe AI (company) — who makes it

Sources