What Is Jev, and Does It Replace LLMs?

Jev is an AI model from the San Francisco startup TypeSafe AI that makes decisions instead of writing text. You send it text or JSON, the structured data your software already passes around, plus questions whose answers you define in advance. It returns one of your answers with a probability attached. That makes Jev a decision model. It reads language like an LLM, the large language model behind ChatGPT and Claude. Unlike an LLM, it cannot write a reply, code or an explanation.

Jev does not replace LLMs. It takes over the small yes-or-no and pick-one calls your software sends to one. A million such decisions a month cost about $34 on Jev, against $8,000 in input alone on Claude Fable 5.1. Diogo Almeida (CEO), Erik Gafni (CTO) and Sasha Sheng (COO) founded TypeSafe in 2024. The company opened early access on 15 September 2026, backed by a $40 million seed round led by DCVC. By 25 September 2026, ten days of outside testing had backed the speed and price but trimmed the accuracy story. Jev's "can't hallucinate" promise covers the format of an answer, not whether the answer is right. An LLM or a person still takes the cases Jev is unsure about.

Why build a decision model?

Many AI calls inside a product are not conversations. They are small judgments. Is this ticket urgent? Does this invoice match the order?

An LLM can answer each one, but it answers by writing. Every answer costs output tokens, the word-pieces models write and bill by. It also costs seconds of waiting, and a reasoning model spends extra tokens thinking first. TypeSafe's launch post puts end-to-end response time at 3 to 329 seconds for frontier models, the largest and most capable LLMs. It puts Jev at 70 to 500 milliseconds.

Scale makes the gap concrete. Take one million decisions a month at 800 input tokens each. The arithmetic is mine, on list prices, and assumes every model counts the text as 800 tokens. Prices come from TypeSafe, Anthropic and, for GPT-5 Nano, Simon Willison. TypeSafe charges only for input, so Jev's input line is its whole bill.

Fable costs 238 times more. Against the cheapest LLMs the gap nearly vanishes, and Jev's case rests on speed and probabilities. I compare hosted model costs more widely in Ollama Cloud vs Claude and GPT, and local ones in my Ollama benchmarks.

The name carries TypeSafe's bet. Jev honours William Stanley Jevons, the 19th-century economist whose paradox holds that cheaper use of a resource can raise total use. TypeSafe expects machine intelligence "to follow a similar path to coal." For an operator, that cuts both ways. As plain arithmetic, a decision that gets 100 times cheaper and runs 1,000 times more often raises the bill tenfold.

How does Jev work?

A Jev request has two parts. The state is the text or JSON you want judged. The questions are what you want to know about it, each under a name you choose. TypeSafe's docs define three question types, and one request can mix all three.

• Noul asks whether a statement is true. It returns one number from 0 to 1, the probability of yes.
• Choice asks which option fits, from a list of up to 255. It returns the pick, a probability for every option and a confidence figure.
• Score places the state on a rubric, a scale of levels you define. It returns a level, a legend, probabilities and a confidence figure.

Each question sees the same state but not the other questions, and all of them run at once. Version 1.13 reads text and JSON only, with no images or audio. The state plus the longest question can run to 32,000 tokens.

The class name explains the design. TypeSafe calls Jev "the first System One model", after the fast, intuitive System 1 thinking Daniel Kahneman popularised in Thinking, Fast and Slow. In his terms, Jev makes the snap judgment and an LLM does the slow, deliberate System 2 thinking. "Decision model" is the press's label, as in TechTarget's headline. TypeSafe's launch post never uses it, and its docs use it on a single page, the AI primer.

Jev's internals are harder to check. TypeSafe says Jev uses a new architecture, a parallel sampler and a training method it calls Reinforcement Learning for Calibrated Decisions (RLCD). According to TypeSafe, Jev outputs all its probabilities in one parallel pass. An LLM instead writes one token after another, each depending on the last. Almeida, a primary author of OpenAI's InstructGPT paper, told TechCrunch that TypeSafe trains Jev only on synthetic data it makes itself. TypeSafe has published no RLCD paper, base model or parameter count.

Is Jev an LLM?

Jev is not an LLM in the everyday sense, because it cannot write text and only returns answers from options you define. TypeSafe's launch post says Jev is "neither small nor an LLM." TechCrunch reports that outside observers suspect an open-weight LLM underneath, one whose model files anyone can download. A preprint by Sun, Xu, Shi and Yang found hosted Jev failing the same way as two open-weight Jev-like models.

My read is a language model with a trained decision layer on top, which TypeSafe has not confirmed. If that read is right, rivals can copy Jev more easily, which matters for its price and for lock-in. Within a week, developers had rebuilt its interface on open models, such as Kev on Qwen3.5.

How does Jev compare to LLMs?

Jev swaps an LLM's open-ended writing for one answer from your options, returned in milliseconds and billed on input only.

Speed and price are where Jev wins. Accuracy needs more care. TypeSafe's homepage promises "193.6x Faster, 444.6x Cheaper," figures its launch post calls "on the higher end of real world gains." They come from workflow evals, test tasks TypeSafe wrote itself. TypeSafe points to those evals but never says which model each figure compares against. Software engineer Pere Pages worked back from the evals and found two different baselines. The speed figure matches Claude Sonnet 5, and the cost figure matches Claude Opus 5.

The evals score agreement with the averaged answers of GPT-6 Astra and Claude Fable 5.1, not correctness. There Jev scores 67.8% against 67.9% for GPT-5.6 Terra, while running about 25 times faster and 76 times cheaper. On invoice processing, though, Jev scores 61.8% against Terra's 74.7%.

The largest independent test is less flattering. Hazem Ibrahim and Yasir Zaki ran Jev against 19 LLMs on 18 social-science labelling tasks, three of them pilots, with 7,977 items. Their preprint is not yet peer-reviewed. On 14 of the 15 evaluation tasks, Jev trailed the best of those LLMs, picked separately for each task. The median gap was 11.6 macro-F1 points, and it was statistically significant on 12 tasks. Macro-F1 is an accuracy score that weights every category equally. Jev's measured cost ran a median 44 times lower. Sending only Jev's low-confidence items to an LLM matched or beat the LLM alone, at a quarter to half of its cost.

Can your LLM already do this?

Structured output is an LLM feature that forces the reply into a fixed form you supply, called a JSON schema. OpenAI's docs say Structured Outputs "ensures the model will always generate responses that adhere to your supplied JSON Schema." A shape guarantee on its own is not new.

Engineer Sean Goedecke explains the real difference. An LLM still writes its structured answer one piece at a time, while Jev "can produce answers to many questions in parallel." His workaround pre-fills the reply so a small open model writes one allowed token. It ran 2x to 3x faster than ordinary structured output. He "suspect[s] Jev does not have a substantial technical moat," yet expects Jev to beat most do-it-yourself copies.

A classifier is a model trained on your labelled examples to sort inputs into fixed categories. Laurie Voss, head of developer relations at Arize, told TechTarget that "Jev doesn't require any pre-training to do that, just a prompt." Once you hold a few hundred labelled past decisions, test Jev, a trained classifier and a cheap LLM on the same set.

jev-phishing-bench shows why that test matters. On 2,000 emails with LLM-written bodies, one broad phishing question gave Jev 62.6% against 81.3% for Claude Haiku 4.5. The author asked five narrow questions and combined them with a logistic regression, a simple weighted formula, fitted on 1,000 labelled emails. On the other 1,000, Jev reached 95.0% and Haiku 93.2%, which the author calls a statistical tie. A two-line text-matching rule scored 91.8% on the same data. Ask narrow questions, and check the simplest baseline first.

What is Jev actually good for?

Jev fits high-volume, narrow judgments inside software, such as routing, triage, labelling and safety checks on agent actions. TypeSafe's launch post names no customers. The evidence comes from builders and researchers, graded here by who measured it.

Pricogni is the clearest lesson for a small company. Its developer designed an LLM judge in June for 9,081 uncertain product matches, then deferred it. "A frontier model at $15 per million output tokens made it not worth building," he wrote. Jev judged all 9,081 pairs in 13 minutes 22 seconds for $0.32. His code rejected 49% of the matches, confirmed 21% and left 30% in the human queue. He judged 48 of 50 hand-checked verdicts sound "by my reading," which he says "is not ground truth." Nothing downstream acts on them yet. The price, not the accuracy, made a shelved feature worth building.

Decisions about people deserve the most caution. Simon Willison called Jev "great for anything that can be expressed as a classification task." He also wrote that "concerns about bias should be front and center" and hopes nobody uses Jev to rank job applicants. A bare probability with no explanation is hard to defend to a rejected candidate.

When is an LLM still better?

Use an LLM whenever the output is words, such as a reply, a summary or an explanation for a customer or auditor. LLMs also win on open-ended reasoning, arithmetic, date logic and input with images or audio, which Jev cannot read. Jev also "cannot extract a value from a document," as Unstract puts it. Unstract's pipeline lets simple code find candidate values and Jev pick the right one. In Unstract's synthetic test, Jev scored a planted wrong amount in an insurance claim at 0.02, close to a certain no. Ten checks on one claim cost about $0.00005, cheap enough to check every field instead of sampling.

Keep a person where a person is accountable, through a human approval step in the workflow. Ned Bellavance, founder of Ned in the Cloud, told TechTarget that "for human-in-the-loop workflows, you're still going to be using an LLM."

Where does Jev fit in agents?

An agent harness is the code around a model that runs its loop, calls its tools and checks its permissions. Inside that harness, many calls are small multiple-choice decisions. Which tool comes next? Should the agent continue, retry or stop? Is this shell command safe to run? Those are Jev-shaped questions, while the LLM keeps the planning and the writing.

Pranit Sharma benchmarked Jev on the command-safety classifier inside Vercel's fx tool. He reported it was "~5-18x faster and more accurate than gpt-5.6-luna, our current top choice." TechCrunch reported that as Vercel replacing Luna, which goes further than his post.

The safest design routes on confidence. Let Jev act when it is confident, and send the unsure middle to an LLM or a person. Ibrahim and Zaki's routing result backs this hybrid. Put that rule in the harness, next to its permission checks and eval gates.

Can Jev hallucinate?

Jev cannot invent an answer outside the options you give it, but it can still pick the wrong one. TypeSafe's homepage FAQ says so itself. Jev "guarantees the shape of its answers, not that every decision is correct." Its launch post still says Jev "can't hallucinate," then concedes its 0% figure "is not empirical."

The Sun, Xu, Shi and Yang preprint shows wrong answers hiding inside valid options. The authors kept each option's meaning and swapped which name it carried, using loaded names like yes/no. Jev's AUC fell from 0.81 to 0.58. AUC measures how well a model's scores rank true cases above false ones, and 0.5 is a coin flip. Answers flipped 24 times more often than Jev's run-to-run noise, while invalid answers stayed at 0%. Neutral names such as 0/1 showed little effect, so name options neutrally.

What do we not know yet?

TypeSafe's own jaggedness page for Jev 1.13 lists nine known failure modes. "Jev is not a calculator," it says, and Jev reads dates "as text, not as ordered quantities." Accuracy "falls as the state grows with content unrelated to the decision." Content written to steer the model "can move the answer." That makes prompt injection, hostile text planted in the input to hijack a model, a live risk.

Calibration matters most and has the least proof. A calibrated model's 90% answers turn out right about 90% of the time, which is the whole reason to route on its confidence. Ibrahim and Zaki found Jev better calibrated than the self-reported confidence of 16 of their 19 LLMs, though three frontier models beat it. On one task, empathy in peer-support conversations, Jev reported high confidence while scoring near chance.

A separate test by GitHub user scienthoon ran 4,621 requests. Jev came out close to calibrated on three familiar public benchmarks, which the author flags as likely training data. On 900 questions about 300 invented support tickets, the average gap between Jev's stated probability and its hit rate more than tripled, to 0.107. Score answers ran clearly overconfident, Choice answers slightly so, and yes-or-no answers underconfident. Your own tickets and invoices are private, so they look more like the invented set than the benchmarks. Recalibrate on a few hundred of your own labelled examples before you route on Jev's scores.

TypeSafe's homepage says it serves Jev profitably at current prices, but its launch post concedes it "can't prove [the price] isn't subsidized." Jev is proprietary, with no published weights or self-hosting option. Voss expects "the major model labs will release decision models very, very quickly," which would test both the price and the moat.

Vercel said nearly 13% of paid teams on its AI Gateway used Jev within 24 hours. Jev was free on the gateway until 25 September, TechTarget reported, so that figure counts trials rather than paying customers.

Early access also carries plain vendor risk. TypeSafe processes data only in the US for now, HiPeople reports. Its docs warn that rate limits "can change without notice" during early access. TechCrunch reported that launch demand briefly ran TypeSafe out of serving capacity. Put these questions to TypeSafe, or to whoever resells Jev to you, before Jev makes live decisions.

• Where is our data processed and stored, and is it used for training?
• What uptime, rate limits and price will you commit to in writing?
• How will we hear before a new Jev version changes our answers?
• Can you show calibration results on data like ours?
• If we leave, what do we keep besides our own labelled data?

What should you do next?

Answer four questions about the decision itself first.

• Can the decision be phrased as yes or no, pick one, or a score?
• Do you have labelled past decisions to test against?
• Does the decision affect a person's job, money or access? If yes, keep a person on every call. The EU AI Act already lists credit scoring, life and health insurance pricing, and candidate screening as high-risk uses.
• What happens to the bill if cheap decisions multiply your volume?

Then pick one high-volume decision your team already makes by hand or with an LLM. Break it into three to five narrow questions with neutral option names, such as 0 and 1. Test Jev and a cheap LLM on a few hundred past decisions. You can start today through Vercel's AI Gateway or DigitalOcean without joining TypeSafe's waitlist. LangChain, Langfuse and Cloudflare already support it.

Set one cut-off to act and one to reject. Pricogni confirmed at 0.8 or above and rejected at 0.2 or below. Everything between goes to an LLM or a person. Go live only on the band where Jev is right at least as often as your current process. Keep the labelled data you build. It lets you recalibrate, switch vendors or train your own classifier when the market moves.

Jev's lasting idea is the price point more than the model. A judgment that costs a fraction of a cent and returns in under a second changes the design question. You stop asking whether you can afford to consult a model. You start asking how many narrow questions to ask it.