# TypeSafe's Jev gives up text to sell probabilities at $0.042 per million tokens URL: https://www.thedeepfeed.ai/posts/2026-09-16-typesafe-jev-system-one/ Category: Models Published: 2026-09-16 Author: the-deep-feed Tags: typesafe, jev, decision-models, calibration, pricing, structured-outputs Kind: news > TypeSafe's Jev answers typed questions with probabilities instead of generated text, which lets it charge $0.042 per million input tokens with free output, but its headline speed and cost multiples come from four workflows the company built itself. ## TL;DR - **Jev** takes a state and a set of typed questions and returns probabilities in one parallel pass, priced at **$0.042 per million input tokens** with output free, about 238 times below the input rate TypeSafe cites for Fable 5.1. - TypeSafe's **193.6x faster** and **444.6x cheaper** figures come from four in-house workflows scored against the average of GPT-6 Astra and Fable 5.1, and the company declined to publish any public benchmark. - By the evening of September 16 the **Hacker News** thread had **1,784 points**, and founder **Diogo Almeida** agreed with a commenter who called Jev a zero-shot classifier: "exactly right!" On September 15, **TypeSafe AI** came out of stealth with [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev), a model that never writes a sentence. A developer sends it a block of state, such as a support ticket or a JSON record, plus a set of typed questions, and Jev returns a probability for every allowed answer in one pass. Input costs $0.042 per million tokens, which TypeSafe also quotes as $42 per billion. Output is free. The company calls this a "System One" model, and the label carries most of the argument: Jev is fast and cheap because it has given up the one thing chat models are built to do, which is generate text. What the launch did not settle is whether its probabilities can be trusted, and the evidence TypeSafe offered for that comes from tests it designed itself. # A model that only answers multiple choice TypeSafe's founder, Diogo Almeida, [writes](https://typesafe.ai/blog/introducing-system-one-models-and-jev) that at OpenAI he "helped build the methods that made language models useful at following instructions," work that became the research behind ChatGPT. Jev is his argument that the same recipe is wrong for automation. The launch post sets the two side by side. Language models are trained with reinforcement learning from human feedback (RLHF) or with verifiable rewards (RLVR), and they sample sequentially, "one token at a time, each conditioned on the last." Jev is trained with what TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD), and it "generates all outputs in a single query." The aim of calibration is that answers given 0.8 probability turn out right about 80% of the time. The trade is enforced at the interface. Almeida told one Hacker News commenter that "strings (and all sequential data structures) are not allowed at all," which is [how TypeSafe computes every output in parallel](https://news.ycombinator.com/item?id=49719122) and why it does not bill for output tokens. The [docs](https://docs.typesafe.ai/concepts/system-one) describe three question types, called primitives: a *noul* (from Bernoulli) returns the probability that a statement is true, a *choice* picks one option from a set and returns the full distribution, and a *score* places the state on an ordered rubric. Choice questions accept up to 255 options. TypeSafe insists this is not a cut-down chatbot. Asked in its launch FAQ whether Jev is "just a smaller LLM," the company answers: "Jev is neither small nor an LLM, hence being off the intelligence Pareto curve." Its [AI primer](https://docs.typesafe.ai/introduction/machine-learning-primer) is more specific about lineage, presenting RLCD as a third way to adapt "pretrained language models," after RLHF and RLVR. Almeida's own gloss on Hacker News splits the difference: Jev is ["technically not a language model (it doesn't generate language)."](https://news.ycombinator.com/item?id=49718437) TypeSafe has published no architecture, parameter count or paper. # The request, the price and the waitlist Every call goes to one endpoint, `POST https://api.typesafe.ai/v1/systemone`, with a `state`, a `model` (the docs use `jev-latest`) and a map of named questions. The [API reference](https://docs.typesafe.ai/api) uses a support ticket as its example: ```json { "state": "Help! My payouts have been failing for 3 days.", "model": "jev-latest", "questions": { "department": { "type": "choice", "instructions": "Which team should handle this?", "criteria": { "billing": "Payments, invoicing, refunds", "technical": "Bugs, outages, integrations", "sales": "Pricing, upgrades, new accounts" } } } } ``` The answer comes back under the same key: `"choice": "technical"`, a probability for each of the three teams, and a `confidence` of 0.82 derived from that distribution. Several questions of different types can ride on one state and are answered together. A noul answer carries only its probability. The price is the headline. TypeSafe's comparison table puts frontier input tokens at $0.20 to $10 per million, with output about five times dearer. [Jev's homepage](https://typesafe.ai/) puts its rate at 238 times below the input price of Anthropic's Claude Fable 5.1, which works out to about $10 per million, the same input rate OpenAI charges for [GPT-6 Astra](/posts/2026-09-04-gpt-6-astra-launch/). The launch post's Doom demo shows what the price means in practice: sending Jev 10 queries a second costs about $7 an hour. TypeSafe concedes that it "can't prove it isn't subsidized" and says it expects the price to fall. Access is narrow. Jev is in early access, with developers admitted from a waitlist, and the service runs from the US West Coast. The company announced a [$40 million seed round led by DCVC](https://finance.yahoo.com/technology/ai/articles/typesafe-ai-emerges-stealth-40m-190000776.html) the same day. # The receipts, as TypeSafe reports them Every performance number in the launch is vendor-reported. The homepage leads with "193.6x Faster, 444.6x Cheaper." The launch post traces those figures to four workflow evaluations (security alerts, agent-trace review, invoice processing and customer service) in which the reference answer is the average of GPT-6 Astra's and Fable 5.1's responses, not human labels. TypeSafe discloses three caveats of its own: the workflows were written by its model-capabilities team, "so some bias could exist"; the competing models ran through TypeSafe's own wrapper; and the multiples are "on the higher end of real world gains." The company also chose to publish nothing on public benchmarks. Its FAQ asks users to "put no weight on public benchmarks" and to build their own evaluations instead. The zero-hallucination chart carries a footnote that "our number is not empirical," because a schema match is guaranteed by design. The speed claims also vary by document. The blog quotes end-to-end latency of 70 to 500 milliseconds, "40x-200x faster." The [Business Wire release](https://finance.yahoo.com/technology/ai/articles/typesafe-ai-emerges-stealth-40m-190000776.html) says "less than 100 milliseconds" and "up to 100 times faster and less expensive." Almeida's launch post on X gave a third pair of ranges: > After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? > > I've spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev > > • 20-200x faster > • 40-400x cheaper (w/ output tokens free) > • Frontier composable intelligence optimized for decisions > > — [@CompleteSkeptic](https://x.com/CompleteSkeptic/status/2099925682726002904), Sep 15, 2026 # The first day's verdict: a classifier, and a large one By 5 p.m. Pacific on September 16, the [Hacker News thread](https://news.ycombinator.com/item?id=49717558) had 1,784 points and 472 comments, according to an archived copy of the page. The most-replied comment, from user jacobgold, accepted the product and disputed the framing: Jev "can only generate structured output," and it "can't emit an invalid type, but it can still emit a completely wrong valid value." Almeida, posting as CompleteSkeptic, [answered](https://news.ycombinator.com/item?id=49718767) that this "is likely true of all ML," adding that he would not say "a random forest 'hallucinates' in the way LLMs do." Elsewhere he [conceded](https://news.ycombinator.com/item?id=49718780) that the models can be "confidently wrong." The description that stuck was older than the branding. User petesergeant called Jev ["basically a zero-shot classifier"](https://news.ycombinator.com/item?id=49718490) that classifies raw or structured text "as accurately (they claim) as a frontier-level LLM." Almeida replied ["exactly right!"](https://news.ycombinator.com/item?id=49718727) Another commenter, bregmandiv, [noted](https://news.ycombinator.com/item?id=49719883) that encoder models already delivered probabilities without hallucination and suggested that what was new was letting a caller define the output shape at request time. A third, mortsnort, [quoted the FAQ](https://news.ycombinator.com/item?id=49722331) against the primer and asked how a post-trained language model is "not an LLM." On X, the first outside numbers and the first backlash arrived within a day: > We benchmarked fx auto mode (safety) classifier with @typesafeai's Jev. > > tl;dr: ~5-18x faster and more accurate than 𝚐𝚙𝚝-𝟻.𝟼-𝚕𝚞𝚗𝚊, our current top choice > > — [@fazxes](https://x.com/fazxes/status/2100300097695232164), Sep 16, 2026 > 12 million views for a JSON classifier? > > Yeah, we're in a bubble > > — [@NielsRogge](https://x.com/NielsRogge/status/2100114968460820986), Sep 16, 2026 Both camps are describing the same product. The open question is the one word TypeSafe chose as the name of its training method. Zero-shot classification at $0.042 per million tokens is a pricing story; calibrated classification at that price, on questions defined by the caller and close to frontier knowledge, would be a new kind of model. TypeSafe has asked developers to judge that on their own data, and until outside evaluations exist, its own four workflows are the only evidence on record. ## Sources - [TypeSafe AI — Introducing System One Models & Jev (Sep 15, 2026)](https://typesafe.ai/blog/introducing-system-one-models-and-jev) - [TypeSafe docs — API reference, POST /v1/systemone (archived Sep 15, 2026)](https://docs.typesafe.ai/api) - [TypeSafe docs — System One and the AI primer (archived Sep 15, 2026)](https://docs.typesafe.ai/concepts/system-one) - [Business Wire via Yahoo Finance — TypeSafe AI emerges from stealth with $40M (Sep 15, 2026)](https://finance.yahoo.com/technology/ai/articles/typesafe-ai-emerges-stealth-40m-190000776.html) - [Hacker News — Introducing System One Models and Jev (Sep 15, 2026)](https://news.ycombinator.com/item?id=49717558) - [@CompleteSkeptic on X — Diogo Almeida announces Jev (Sep 15, 2026)](https://x.com/CompleteSkeptic/status/2099925682726002904) --- Canonical: https://www.thedeepfeed.ai/posts/2026-09-16-typesafe-jev-system-one/ Site: https://www.thedeepfeed.ai Full corpus: https://www.thedeepfeed.ai/llms-full.txt