Jev by TypeSafe AI: what it is and how it differs from a traditional LLM

TypeSafe AI just launched into early access a model that breaks the very definition of "LLM": Jev doesn't generate a single word of text. It receives a question with predefined answers and returns a decision with a probability attached — nothing more. That apparent limitation is, according to its creators, exactly why it can be up to 100 times cheaper than GPT-5.6 or Claude Opus 5 for certain tasks.

What is Jev?

Jev is the first model in what TypeSafe AI calls the "System One" class — a name drawn directly from Daniel Kahneman's distinction between fast, intuitive thinking (System 1) and slow, deliberate thinking (System 2). It's built by Diego Almeida, who worked at OpenAI as a co-creator of ChatGPT and of RLHF techniques before founding TypeSafe. The name "Jev" references economist William Stanley Jevons, known for the paradox bearing his name about how cheaper access to a resource drives up its total consumption — a declared bet that much cheaper AI decisions will multiply how much AI gets used, not replace generative LLMs.

The core difference: it decides, it doesn't write

A traditional LLM — GPT-5.6, Claude Opus 5, Gemini 3 — generates text token by token, autoregressively: each word depends on the ones before it, which is why a long answer takes seconds. Jev works differently: it uses a parallel sampler that answers every question in a single pass, without generating any free-form text at all. Instead of "writing a response," the developer defines the space of valid answers up front — just three question types exist: Choice (pick among named options), Score (a value on an ordered scale), and Noul (yes/no, expressed as a probability between 0 and 1). Jev can never return anything outside that schema — not because it's "well-trained" to avoid it, but because it's mathematically impossible given how the model is built.

Training is different too: instead of RLHF (reinforcement learning from human feedback, optimized so the text appeals to a human), TypeSafe uses what it calls RLCD — Reinforcement Learning for Calibrated Decisions. The goal isn't for the answer to "sound good," but for the probability the model reports to genuinely match its real hit rate. If Jev says "82% probability," the idea is that in practice it's right about 82% of the time — a property a normal generative LLM doesn't come close to guaranteeing.

The numbers — with the caveat that they come from the vendor itself

TypeSafe reports striking figures: between ~40x and ~200x faster than frontier models on comparable tasks, and between ~40x and ~400x cheaper, with latencies of 70-500 milliseconds versus the 3-329 seconds typical of a generative LLM, at $0.042 per million input tokens (output isn't billed, since there's no text to generate). None of these figures are independently verified — they're TypeSafe's own internal benchmarks, and the very sources we consulted for this article point that out explicitly.

What Jev can't do

The list of limits matters as much as the capabilities. Jev doesn't generate text, code, summaries, or explanations — there's no way to ask it "explain why" for a decision, only the probability. It doesn't process images or audio. It doesn't do multi-step reasoning: TypeSafe is explicit that you need to route that to a separate reasoning model. It only works with text input (mostly English), with a 64,000-token context budget and a maximum of 255 options per Choice-type question. It's API-only (its own POST /v1/systemone endpoint, not compatible with the standard chat interface), hosted in the US with no documented EU region, and is in early access with a waitlist.

What this means for a company already using or buying AI

Jev doesn't compete with your chatbot or the agent that drafts emails or code — that's not what it's for. Where it genuinely makes sense is as cheap infrastructure for repetitive, bounded decisions at scale: routing support tickets to the right team before an expensive LLM ever touches them, classifying sentiment across millions of reviews, filtering content or detecting jailbreak attempts on another model's output, or making real-time decisions (dozens per second) in a transactional flow. The pattern TypeSafe is proposing is clear: a fast model decides and routes, a generative model writes when text is actually needed, and a person steps in when confidence falls short. For a company in Guatemala or Latin America deciding where to put its AI budget, the right question isn't "do I replace my LLM with Jev?" but "do I have a high-volume classification process where paying $0.18 per case with Claude Opus 5 doesn't make sense when the decision is binary or multiple-choice?" If the answer is yes, it's worth a look — as a complement, not a replacement.