TypeSafe AI's Jev Model Skips Text Generation and Charges $0 for Output Tokens

TL;DR

TypeSafe AI came out of stealth on September 16 with a $40M seed led by DCVC and a model called Jev that gives up text generation entirely. You send it state plus a schema of typed questions and get one answer per question, each with a probability distribution and a confidence, in 70 to 500 milliseconds. Input costs $0.042 per million tokens; output is free, because there is no autoregressive decode to bill. On TypeSafe's own four-workflow eval it matches Sonnet 5 and GPT-5.6 Terra on accuracy at 1/76th to 1/290th the cost per case. The catches: the eval is self-run with LLM-derived labels, the model reads instructions literally, cannot count or compare dates, and prompt injection still moves its answers.

If you run classification, routing, or guardrails at volume, read on. If you wanted a cheaper chatbot, close the tab.

What actually shipped

"Models have been superhuman at chat for years, so where is all the automation?" That is the first line of the launch post, and CEO Diogo Almeida calls it his driving question for the last four years. TypeSafe was founded after he left OpenAI, together with Erik Gafni and Sasha Sheng. TypeSafe's docs credit Almeida as a co-inventor of RLHF, the technique behind InstructGPT and ChatGPT, which makes the pitch pointed: the person who helped teach models to please humans now says production automation needs a different objective. Two years in stealth produced a model named after the Jevons paradox: a 100x cheaper decision does not shrink your bill, it multiplies the number of decisions you make.

The shipped artifact is Jev 1.13 (model id jev-1.13.0, alias jev-latest), served from a single endpoint, POST /v1/systemone, behind an early-access console at console.typesafe.ai. SDKs: Python and TypeScript. Context is 64k tokens per request, with the state plus the longest single question capped at 32k. Rate limits are 250,000 tokens per second and 1,200 requests per minute, "adjusting dynamically" while "upcoming large GPU deals land." Input is text only; images, audio, and video must be pre-processed into text first.

TypeSafe calls the category a System One model, after Daniel Kahneman's fast, intuitive System 1. Almeida's version: "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out."

Three primitives and nothing else

A Jev request has three fields: state (a string, a JSON object, or an array of text), model, and questions, a map from ids you choose to typed questions. There are exactly three question types:

• Choice: pick one option from a set you define, up to 255 options. Returns the winning choice, a probabilities map summing to 1, and a confidence from 0 to 1.
• Score: rate the state against an ordered rubric of 2 to 10 described levels. Returns a probability per level and a score that is the probability-weighted position, so 0.70 on level 1 and 0.30 on level 2 gives 1.30.
• Noul: a single yes/no statement. Returns one number, the probability the answer is yes. No confidence field, because the number already is one.

That is the entire output vocabulary: no string type, no free-form JSON, no "explain your reasoning" field. TypeSafe's 0% structural error rate means exactly this: the model cannot emit a key you did not declare or an option you did not list, because it never emits text at all. Whether it picked the right option is a separate question, and the company says so.

Why the output is free

An LLM answers a multiple-choice question one token at a time, each conditioned on the last, usually with a paragraph around the answer. That sequential decode is why output tokens cost several times more than input everywhere, and why a routing call that should take a blink takes seconds.

Jev ingests the state once and, per the docs, "evaluates every question against it in parallel." AI News reports the 255-option Choice is handled by two-stage independent scoring of the options, not by generating the option name. No decode loop means no output to meter, so TypeSafe charges nothing for it. The mechanism beyond that is unpublished, which is normal for a lab's first model and worth remembering when reading the speed claims.

Think of a multiple-choice exam. An LLM writes an essay for every question and you fish the letter out of the essay. Jev is the Scantron machine: it reads every bubble in one pass and hands back a probability per bubble. The Scantron cannot write you an essay, and that is the whole product.

The launch demo was a Doom bot querying the model 10 times a second for about $7 an hour, which The Register clocked at 0.114 seconds per decision against 8.566 seconds for GPT-5.6 Terra on the same game state. It is the most expensive way to play Doom since 1993, and the only one where the player is a classifier.

RLCD: rewarded for knowing how unsure it is

TypeSafe's primer lays out three post-training paths: RLHF, which optimizes for responses people prefer and which the company blames for sycophancy and "confident-sounding hallucinations"; RLVR, which built reasoning models that are strong at math but slow and expensive; and its own Reinforcement Learning for Calibrated Decisions, where the model never produces text and its probabilities are "optimized against outcomes to reflect uncertainty."

Calibrated has the strict meaning here: across many predictions, outcomes assigned 0.2 should happen about 20% of the time and outcomes assigned 0.8 about 80%, a property of groups of predictions, never of any single answer. A good weather forecaster is calibrated in exactly this sense: it rains on about seven of every ten "70% chance" days, and that tells you nothing about tomorrow.

The confidence value on Choice and Score answers is derived from the shape of the distribution (peaked is high, flat is low), with the full distribution returned if you want your own measure. The recommended pattern is three bands: act on high confidence, confirm or flag on medium, route low to a person or a reasoning model.

The benchmark TypeSafe ran on itself

TypeSafe published a workflow eval across four tasks (security incident triage, agent trace observability, invoice processing, customer service) and nine models. Each task is decomposed into Noul, Choice, and Score questions with the control flow in code, and every model runs the same decomposed workflow. Means across the four, equally weighted:

Read the fine print before repeating the numbers. The reference labels are not human: they come from GPT-6 Astra and Claude Fable 5.1 at high reasoning settings, and the eval page says so outright: "we assume that the code is correct, and measure against the current smartest large models." Every other model ran at provider defaults, not its best thinking budget, no per-workflow case counts are published, and no independent reproduction exists yet. The "193.6x faster, 444.6x cheaper" on TypeSafe's homepage comes from this same table, and the launch post adds that "we expect that these are on the higher end of real world gains," a caveat that did not survive into most of the coverage. The defensible reading: on decomposed decision workflows with LLM-judged labels, Jev lands in the mid-tier accuracy band for two to three orders of magnitude less money and time. For a first model, that is still a remarkable place to land.

What "cannot hallucinate" actually means

The guarantee is structural, not semantic. Jev cannot return malformed output, an undeclared key, or an option outside your list. It can absolutely put 0.92 on the wrong option, and it will never tell you why, because there is no rationale channel. A Choice with no fitting option does not abstain, it picks the nearest one, which is why the docs tell you to add an "other" or "none of the above" option whenever the list might not cover every input. Where a reviewer needs the reasoning, that is a hybrid-pipeline requirement, not a footnote.

To its credit, TypeSafe ships a jaggedness page that reads like the failure log most vendors keep private. Jev 1.13 answers "the question you wrote, not the one you meant," so scoping words and negations are read literally. It does not count reliably or do date arithmetic. Accuracy drops as irrelevant detail piles into the state, and double negatives or multi-hop indirection cost more. Invariants you would assume, like a Noul and an equivalent yes/no Choice agreeing, are not guaranteed. And the line that matters for guardrail builders: "State is data, and jev-1.13 does not treat it as hostile by default." Injected instructions inside the state can move the answer. TypeSafe expects to improve this; until then, a Jev-based prompt-injection filter is itself injectable, so test that before you ship it.

Where it slots in next to an LLM

System 1 takes the fast, bounded, high-volume judgments: triage, routing, guardrails, scoring. System 2, your frontier LLM, takes anything that must produce prose, code, or an explanation.

Three workload shapes fall out of that:

• Triage. A support ticket goes in as state; a Choice over every team and a Score over severity come back in one call. Low confidence on the team question means copy a second team or ask a human; the docs' worked example does exactly that at 0.39.
• Guardrails. A Noul asking whether a prompt is a jailbreak attempt, or a tool call is in scope, runs before you wake the expensive model. At 0.4 seconds and $0.0004 a case it can run on every turn. Just remember the jaggedness page: it is a filter that can be talked to.
• Big data. The Jevons bet in numbers: at $42 per billion input tokens, scoring a million 1,000-token records against a 24-topic Choice costs about $42, output included. That is the price at which nobody bothers to sample first.

The handoff is confidence-gated: code acts on the high-confidence answers and forwards the rest, state included, to an LLM that can reason and explain. The same primitives work on the way back out: a Noul over the LLM's draft ("does this reply answer the question", "does it contain personal data") is a 0.4-second inspector on generated output, and TypeSafe's agent-trace workflow in the eval is exactly that shape, deciding whether a person needs to look at a finished agent run. The second launch demo, a Wikiracing bot that picks which link to follow from a page full of them, is the same trick pointed at a DOM. TypeSafe also publishes a System One adapter, a drop-in replacement for its client that runs the same state-and-questions schema against an LLM API, so you can A/B the two or keep a fallback that speaks the same types.

The parts to stay skeptical about

Three things. First, every number above is TypeSafe's, from a harness TypeSafe wrote and labels from models TypeSafe chose; treat cost and latency as reproducible and accuracy as provisional until someone outside runs it. Second, the price. A $40M seed, a rate-limit note about GPU deals still landing, and a model 76x cheaper per case than GPT-5.6 Terra describe a company buying share, and the docs already warn that limits "can change without notice." Pin jev-1.13.0 rather than the alias if you tune thresholds, because the alias moves. Third, The Register's objection stands: a classifier's error rate and a text generator's hallucination rate are different failure surfaces.

None of that changes the builder-level fact: there is now a frontier-lab model whose entire interface is a typed function signature, whose output is free, and whose docs tell you where it breaks. For the decision-shaped majority of an agent loop, that is the right abstraction, and someone finally shipped it.

Key Takeaways

• Jev outputs only three typed primitives, Choice (up to 255 options), Score (2 to 10 levels), and Noul (a yes/no probability), each with a full distribution, so structural output errors are impossible by construction.
• $0.042 per million input tokens, output free; TypeSafe's eval puts it at $0.0004 and 0.4 seconds per case, behind rate limits (250k tokens/second, 1,200 requests/minute) the docs say are still moving.
• On TypeSafe's own four-workflow eval Jev scores 67.8%, level with Sonnet 5 and GPT-5.6 Terra, behind Opus 5 (73.1%) and Sol (74.1%); labels come from GPT-6 Astra and Fable 5.1, not humans, and nobody has reproduced it yet.
• "Cannot hallucinate" means cannot emit an undeclared key. It can still pick the wrong option, gives no rationale, reads literally, cannot count or compare dates, and can be moved by adversarial content in the state.
• Use it as the System 1 layer (triage, routing, guardrails, bulk scoring, and a Noul inspector on LLM output) with confidence-gated handoff to an LLM for anything that must reason or explain.
• TypeSafe raised a $40M seed led by DCVC, was founded by RLHF co-inventor Diogo Almeida after two years in stealth, and named the model after the Jevons paradox: cheaper decisions mean more decisions, not smaller bills.

Sources: TypeSafe AI: Introducing System One Models and Jev, TypeSafe docs: Models, TypeSafe docs: API reference, Choice, Score, Noul, Confidence, AI primer (RLCD), Jev 1.13 jaggedness, TypeSafe workflow evals, system-one-adapter-python, The Register, AI News, AIwire: TypeSafe AI emerges from stealth with $40M, Tech Startups