Jev by TypeSafe: a decision model that isn't an LLM

On September 15, TypeSafe AI unveiled a new class of models it calls "System One", along with the first of them, Jev (announcement). The class name nods to Daniel Kahneman's fast, intuitive System 1; the model is named after William Stanley Jevons. Founder Diogo Almeida previously worked on instruction-following methods at OpenAI.

The short version: this is not a chatbot. It is a fast, cheap decision layer meant to sit inside your code.

How it works

You send a state (a string, a JSON object or an array of text) and ask several predefined questions about it at once. There are three question types (docs):

Choice picks from up to 255 options and returns a probability for each. Score places the input on an ordered scale. Noul (short for Bernoulli) returns the probability that a yes/no statement is true.

Jev produces no explanations, code or prose. On Hacker News, Almeida agreed with a commenter who described it as essentially a "zero-shot classifier" (HN thread).

Typical uses: routing tickets, classification, scoring, filtering, or quickly checking an LLM agent's output. Your own code handles the logic and combines the answers.

Pricing, versions and context

According to the official models page, input costs $0.042 per million tokens ($42 per billion), and output is free (Models).

The newest release is Jev 1.14, available since October 14, 2026, behind the jev-preview alias. The stable jev-latest alias still points to 1.13.

Context is 64k tokens per request and 32k per question. Sending the TypeSafe-Danger-Zone-Long-Context header raises both to 262k, though the docs warn this is slower and less accurate. Input is text only.

There is a discrepancy here. OpenRouter's model page lists Jev 1.13 with a 262K context and "text and images" input, and shows a P50 latency of roughly 0.1 s (0.13 s when we checked) (OpenRouter). TypeSafe's own docs mention text only, and offer 262k only via the special header. On OpenRouter, Jev runs through a separate Decisions API rather than the usual chat endpoint (OpenRouter tutorial).

What "can't hallucinate" actually means

TypeSafe's own blog says the 0% on its hallucination chart "is not empirical". It follows from the fact that schema matching is guaranteed.

In plain terms, Jev cannot return a value outside the options you define. A well-formed answer can still be wrong. Fractional's analysis notes that TypeSafe's FAQ says the same: the shape of the answer is guaranteed, not its correctness. In one community audit, Jev put a probability of 1.00 on a wrong answer (Fractional).

Why it matters: this guarantee isn't new. OpenAI has offered structured outputs via constrained decoding since August 2024, and other providers followed. In NavyaAI's test, strict JSON mode produced zero unparseable responses in 600 calls (NavyaAI). Jev's real differences lie elsewhere: native per-option probabilities, parallel questions over one state, input-only billing and speed.

193.6× and 444.6×: the vendor's own figures

The "193.6x faster, 444.6x cheaper" claim on TypeSafe's homepage is the company's own measurement. It comes from four workflows written by its own team, scored against the average of GPT-6 Astra and Fable 5.1 as a reference. TypeSafe itself concedes these are at the higher end of real-world gains and that some bias may exist.

Independent numbers are far more modest. Fractional finds Jev roughly 2–3.6× faster than LLMs configured for decisions and roughly 5–50× cheaper per decision. It calls the 193.6× figure exaggerated, since it benchmarks against slow, expensive configurations.

In NavyaAI's AG News topic-classification test (200 cases), Jev scored 87.5–88% against 70–83% for small LLMs. But it was not cheaper than the cheapest do-it-yourself setup ($18.57 vs. $16.30 per million decisions), and only about 1.7–1.9× faster than it.

Who should use it

Jev is strongest where there are many decision types and no labeled data. It delivers good accuracy without labels.

With a stable taxonomy and a few hundred labeled examples, the picture changes: in tests cited by Fractional, a fine-tuned small classifier such as DistilBERT beat Jev on several tasks. Both analyses recommend recalibrating the probabilities on your own data, and not treating the confidence field as the chance of being right.

Before you wire it in

The architecture and the RLCD training method are not publicly documented. The docs say English is the strongest language, so test on your own content for anything else. Aliases move with new releases, so pin a version in production (for example, jev-1.13.0).

The practical next step: run a few hundred of your own labeled cases through it, compare against a small LLM with structured outputs, then decide. Jev may well win, but that should come from your data, not the marketing page.