Jev

Spec

Attribute	Value
Developer	TypeSafe AI
Released	2026-09-15
Announced	2026-09-15
Context window	unknown
Pricing	$0.042/M input · output free
License	unknown
Availability	early access off a waitlist at typesafe.ai
Three of these rows need their reading stated, because each is unknown or odd
for a different reason
(source):

• Context window is unknown and not a number. One pass reports a request budget of roughly 32,000 tokens, described in TypeSafe's docs as about 150,000 English characters. No pass used the term "context window", and a request budget for a model that makes one parallel forward pass over a state is not obviously the same quantity. The figure is recorded here; the row stays unknown until something names it.
• Pricing carries a free half. TypeSafe states input at $42 per billion tokens (= $0.042/M) and output free, its stated reason being that the architecture makes output too inexpensive to meter (3 passes). No Catalogue id row is present: Jev is in waitlisted early access and appears in no catalogue any pass returned, so spec-check will report it not-listed, which is the correct answer for this model rather than a defect.
• License is unknown, not proprietary. Nothing read names a licence or terms of any kind.

Release Date

2026-09-15, announced together with the System One Models class in a post titled Introducing System One Models & Jev (2 passes). One pass dates the opening of early access 2026-09-16 — recorded in ## Conflicting Reports (source).

What it is — a model that cannot answer you in words

Jev does not generate text. The caller sends structured or natural-language state together with a schema; the model returns a typed value inside that schema — a choice, a score, or a probability with a stated confidence — in one parallel forward pass rather than token-by-token decoding (3 passes).

Property	Value	Passes
Stated jobs	decide, classify, route, score	3
Output shape	typed value + calibrated probability / confidence	3
Decoding	one parallel forward pass, no autoregressive generation	3
Schema guarantee	stated that the model cannot emit an output outside the caller's schema	2
End-to-end latency	70–500 ms	3
Training method	RLCD — Reinforcement Learning for Calibrated Decisions	2
Request budget	~32,000 tokens ≈ 150,000 English characters	1
RLCD is named, not specified. Its stated goal is calibration — a stated 70%
confidence should be correct about 70% of the time. One pass records that the
**reward function, architecture, training procedure and calibration methodology
are all undisclosed** (source).

Benchmarks

2026-10-22 — controller evaluation of Jev 1.13. Good Judgment, Bad Advice tests whether decision signals improve a fixed terminal-agent loop, with rule-based, small-model and placebo controls. The paper distinguishes offline ranking from useful interventions; its abstract establishes neither the author's affiliation nor general performance across Jev deployments. (source)

There are no independent evaluations of Jev at launch (2 passes). What exists is one vendor figure (source):

Measure	Value	Harness	Passes
Accuracy vs "the most expensive frontier models"	within 3 points	TypeSafe's own workflow evaluations	1
Cost per call vs the same	~4,000× less	TypeSafe's own	1
**The launch source contains no benchmark this wiki tracks — not Artificial Analysis, not DeepSWE, not
Terminal-Bench, not any public suite.** Its comparator models are not named.

Two evaluations from outside TypeSafe arrived on 2026-09-24, both from the same HuggingFace Daily Papers snapshot (source):

Measure	Value	Source
Gap to strongest comparator, preference + evidence-grounded factuality	within 3 points, against 16 judges, blinded human adjudication	JEV-as-a-Judge: Accept When Confident, Escalate When Unsure
Cost against that comparator	0.36% of the fee	JEV-as-a-Judge: Accept When Confident, Escalate When Unsure
Frozen confidence cascade	retains 99% of comparator accuracy	JEV-as-a-Judge: Accept When Confident, Escalate When Unsure
LoCoMo LLM-as-a-Judge, System-One-controlled memory	0.777, +11.0% relative; construction 158 s (6.6×), query latency 0.93 s (−36.7%)	Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents
Neither snapshot entry carries an author list or affiliation, and
arxiv.org answers EGRESS_BLOCKED from this run's sandbox, so **whether
either paper is independent of TypeSafe is unknown and is not asserted here**.
The strongest comparator is not named in either.

The September snapshots provide no public measurement of calibration, the property RLCD is stated to optimise. The judge paper's cascade presumes calibrated confidence and reports that JEV's gap to the comparator is concentrated in low-confidence decisions — which is indirect evidence that the confidence signal carries information, and is not a calibration measurement.

Pricing

$0.042 per million input tokens; output free (3 passes). For contrast within this wiki, Fugu Max — the other 2026 release that sells a non-standard object as one model — lists $2/M input · $6/M output, and Claude Opus 5 $10/M input.

The comparison is not like-for-like and this page does not present it as one: Jev returns a value, not a completion, so a "call" here is not the same unit of work. The 4,000× cheaper per call figure above is TypeSafe's own attempt at a like-for-like unit and is unverified.

Use Cases

Stated by TypeSafe, and the list is the product boundary rather than a set of suggestions (3 passes) (source):

• Routing a request to the next step in an automation workflow
• Classifying unstructured state into a caller-defined enum
• Scoring a candidate, a result, or a confidence
• Gating — a yes/no decision with a calibrated probability attached

Anything requiring prose is out of scope by construction. The pitch summarised by one pass — "a frontier-intelligence function call" — is the honest description of the surface.

Compared To

• Fugu Max and Fugu Ultra v2 — Sakana AI's orchestrators, the other 2026 entrants sold as "a model" while not being a single trained network. They move in the opposite direction: Fugu makes one endpoint out of many frontier models and charges frontier-adjacent prices; Jev removes the language model from the decision entirely and charges ~1/50th of Fugu Max's input rate. Both are answers to the same observation — that a lot of production traffic is paying frontier prices for non-frontier work.
• "Small frontier LLMs" — the comparator TypeSafe's own headline names and never identifies. Until it does, every multiplier on this page is a ratio with one side missing.

Conflicting Reports

The cost multiplier against an outside measurement

TypeSafe's own figure is ~4,000× less per call than "the most expensive frontier models" on TypeSafe's own workflow evaluations. JEV-as-a-Judge: Accept When Confident, Escalate When Unsure reports 0.36% of the strongest comparator's fee on a judging task, which is ~278× — a 14× discrepancy. The two are not measuring the same unit: one is TypeSafe's per-call construction against unnamed models, the other a per-task fee against an unnamed state-of-the-art judge. Neither figure is adopted over the other (source) (source).

The speed and cost multipliers

TypeSafe publishes four different answers for the same two quantities across its own surfaces — 193.6×/444.6× (home page), 40–200× (blog), 20–200× / 40–400× (founder's thread), >100×/>200× (the AINews headline carried first-party in this repo's feed). The full table and its reading are on TypeSafe AI. No figure is adopted here (source).

The early-access date

Two passes give 2026-09-15 for the launch and the opening of early access; one pass states early access opened 2026-09-16. The Released and Announced rows carry 2026-09-15 as the better-corroborated reading, and the alternative is recorded rather than discarded (source).

Sources

• sources/blogs/typesafe-2026-09-15-jev-system-one.md
• sources/papers-daily/hf-daily-2026-09-24.md
• AINews — Jev: a "System One Model"
• AI News — ChatGPT pioneer launches Jev model for programmatic logic
• Developers Digest — TypeSafe Jev: the first decision-only model class