Overview

Jev is TypeSafe AI's first public System One model, launched in early access on September 15, 2026. It is an API model for developers who need software to make bounded judgments—such as routing a support request, scoring a record against a rubric, or checking a condition—without parsing generated prose. An application sends text or structured state with predefined questions; Jev returns typed answers and probabilities that the application can handle in code.

The distinction is also its boundary. Jev does not write emails, code, summaries, or arbitrary JSON text. It is a decision component inside a workflow, not a general-purpose chatbot or autonomous agent. The current version documented by TypeSafe is jev-1.13.0; jev-latest and jev-preview pointed to it when this page was checked. TypeSafe's September 15, 2026 launch post described Jev as early access and said developers were being brought off a waitlist; no newer official access-status statement was verified for this review.

Key Features

Choose from defined options — Choice returns a selected option, a probability for each candidate, and confidence, allowing application code to route or rank bounded outcomes.
Score against a rubric — Score evaluates state against ordered criteria and returns a score and probability distribution for a defined scale.
Check a condition — Noul returns a 0–1 value for a yes-or-no judgment; the application decides the action threshold.
Ask in parallel — A single request can mix the three question types against one shared state. TypeSafe says each question is evaluated independently, so broad decisions can be decomposed and recombined in code.

Where Jev Fits

Jev is most useful when the set of possible answers can be defined before the call, the input is already text or structured fields, and downstream code owns the business rules. A support-triage system, for example, could ask one Choice question for department and separate Noul questions for urgency and escalation, then apply its own thresholds. That is a workflow pattern, not a claim that Jev has a packaged help-desk product.

For production decisions, pin jev-1.13.0 if thresholds have been calibrated against that version. TypeSafe notes that the jev-latest alias can move to a new model while application code remains unchanged. The vendor recommends testing on the application's own data, especially outside English.

Pricing & Access

TypeSafe lists Jev at $0.042 per million input tokens, with output tokens free, as of September 20, 2026. Billing is token-based rather than a fixed per-seat subscription. TypeSafe's September 15, 2026 launch post described Jev as early access and said it was bringing developers off a waitlist; the published API price should not be read as proof of immediate account access. No general free tier or self-service entitlement was verified for this page.

The documented limits are 64,000 tokens for a full request, 32,000 for the state plus its longest question, 250,000 tokens per second, and 1,200 requests per minute. TypeSafe warns that rate limits are changing dynamically. These figures are a dated snapshot, not an SLA or a promise of throughput for a new account.

Limitations & Trade-offs

Typed output is not factual correctness. Constraining the answer shape prevents out-of-schema strings, but Jev can still choose the wrong option or assign an unhelpful probability. TypeSafe's own limitations page documents these failure modes.
Exact arithmetic belongs in code. TypeSafe says Jev is unreliable for counting, mathematical operations, date ordering, and numeric precision. Precompute those values and reserve the model for semantic judgments.
Question wording matters. Version 1.13 can interpret instructions literally, struggle with indirection or contradictory criteria, and be steered by adversarial text in the state. Explicit criteria and application-level validation are necessary.
Text in, decisions out. Images, audio, and video must first be converted to text or structured fields. Jev is not trained to generate free-form text; TypeSafe says forcing generation by chaining choices works poorly and is very slow. English is the language where TypeSafe says accuracy is currently best.
Early access limits adoption. A developer must obtain access before directly using TypeSafe's API, and documented rates or limits may change.

TypeSafe publishes favorable speed and cost comparisons for its own System One workflows, but those evaluations are vendor-run and task-specific. This page does not treat them as independently verified performance for every workload. The practical decision is whether typed, probabilistic judgment—not writing or open-ended reasoning—is the bottleneck in the workflow.

Best For

Developers building API-based classification, routing, scoring, or guardrail logic with bounded answers.
Teams that can define clear criteria, test confidence thresholds on their own examples, and keep deterministic rules in code.
Text-first applications where several independent judgments must be made against the same state.

Avoid If

You need generated text, code, explanations, or open-ended extraction instead of a choice from defined possibilities.
The task depends on exact calculations, date comparisons, or multi-step reasoning that ordinary code should perform.
You need image, audio, or video inputs without a preprocessing step, or immediate self-service availability is mandatory.
You cannot evaluate the model's decisions and threshold behavior on representative data before automation.

Sources & Verification

Last verified: 2026-09-20
Evaluation scope: Product role, question types, access, pricing, published limits, and documented version 1.13 failure modes. No independent latency, calibration, or accuracy benchmark was run.
Method: Source-based research from the official product materials below. No hands-on testing was performed for this update.
Official sources:

Jev