News · 2026-09-22
Jev: A System One decision model for agents (not a chat LLM)
TypeSafe AI’s new model returns typed probabilistic decisions for agent routing, scoring, and guardrails rather than free-form text.
What it is
Jev is TypeSafe AI's first public System One model, announced on 15 September 2026 by founder Diogo Almeida and available in early access. It is designed for fast, typed decisions that software and agents can consume directly, rather than for chat, code generation, or open-ended prose.
The model accepts an unstructured state plus bounded, typed questions and returns typed results with probabilities and confidence. Multiple questions can be evaluated against the same state in a single request, each in parallel and in isolation.
Jev is not an autonomous agent and not a replacement for a generative LLM. It is a typed decision layer that can be used within agent workflows: the decision layer routes, scores, and gates, while a generative model remains responsible for planning, explanation, and free-form work.
Core technical characteristics
- Three typed primitives: Choice (select from a defined set, returning choice, probabilities, and confidence), Score (place input on an ordered rubric, with score, probabilities, and confidence), and Noul (a calibrated yes-or-no statement test returning a 0–1 value).
- Parallel evaluation: all questions against one state are answered in a single request, evaluated independently, so added questions barely change response time.
- Calibrated confidence on Choice and Score outputs, intended for thresholding automatic action versus escalation to review.
- Vendor-stated input pricing of $42 per billion input tokens ($0.042/MTok), with output tokens described as too cheap to meter — a vendor claim, not an independently verified cost.
- Vendor-stated end-to-end response times of 70–500ms for System One shaped queries — a vendor claim measured on vendor-run evaluations.
- Schema-guaranteed output structure: outputs conform to the pre-defined schema, so type errors are prevented by construction (vendor design claim).
- Atomic-question design: each question should ask one well-scoped thing; complex judgments are decomposed into separate questions combined in application code.
Why it matters
Many agent steps are bounded classification or routing problems: which tool to call, whether to approve an action, how to score an input, whether a task is complete. A structured decision interface avoids parsing free-form text and lets application code define thresholds for automatic action or escalation.
The surrounding code constrains the model's freedom, which makes decisions easier to compose into reliable systems: classify, route, score, extract, or branch where hand-written rules are too brittle, then verify, guardrail, and monitor around them.
TypeSafe positions this pattern for AI-powered workflows, map-reduce over large datasets, real-time applications with latency budgets, and scoring or verification of LLM outputs — each viable only if decision quality is measured on the workload in question.
Limitations and open questions
Latency, pricing, and efficiency claims are vendor-stated and not independently established. The headline figures of up to 193.6x faster and 444.6x cheaper come from TypeSafe's own workflow evaluations, which use the average of two external models (GPT-6 Astra and Claude Fable 5.1) as the reference answer on workflows authored by TypeSafe's own capabilities team — a design the vendor itself flags as potentially biased.
The “can't hallucinate” framing requires qualification: what is guaranteed by construction is schema conformance (no type errors), not decision correctness. An incomplete state, ambiguous labels, or poorly bounded choices can produce weak or wrong decisions at high speed.
TypeSafe itself notes that its pricing sustainability is unproven and that published speed evaluations were run from its own infrastructure. Independent, standardized evaluations across diverse agent workflows remain limited.
As a newly released system in early access, APIs, models, integrations, and operating guidance may change quickly. Production use should include monitoring, safe fallbacks, and human review for consequential actions.