What is Jev? TypeSafe's new AI model.

The short answer

Jev is an AI model that makes decisions instead of writing. You give it some information and a question with fixed answer choices, and it picks an answer and tells you how sure it is. It cannot write a sentence.

Ask ChatGPT whether an email is a refund request and it writes back a paragraph. Ask Jev the same question with the choices "refund, billing, or cancel" and it returns something like "refund, 91% sure." One is an essay answer. The other is a multiple-choice answer with a confidence score attached.

That is less than a chatbot does, on purpose. What Jev trades away in flexibility it gets back in speed and price. TypeSafe AI, the company that makes it, reports answers in 70 to 500 milliseconds and charges $0.042 per million input tokens (a token is roughly three-quarters of a word), with nothing charged for the answer.

Jev is not an app you chat with. Developers build it into their own software, where it makes the same small call thousands of times a day: sorting support tickets, flagging risky ad copy, scoring sales leads. TypeSafe launched it in early access on September 15, 2026, and opened it to everyone on September 20.

How Jev compares to ChatGPT, Claude, and Claude Code

ChatGPT and Claude write, and Claude Code writes and runs software. Jev sorts, which puts it closer to the spam filter in your inbox than to any of them.

ChatGPT and Claude are chat assistants built on large language models. They read what you type and write a reply one word at a time, so they can explain, draft, summarize, and reason through a problem where you can see it. That flexibility costs time and money on every question, and it leaves you with a paragraph that a person, or more code, has to turn into a decision.

Claude Code is Anthropic's coding agent. You give it a task and it reads your files, writes code, runs commands, and checks its own work over many steps. Jev can do none of that, because it cannot write. The two can work together, though: Alex Volkov used a Jev plugin to shrink a long Claude session from nearly 1M tokens to 86K in about a second, a demo covered below.

A spam filter is the closest thing most people already use. It sorts one kind of message into fixed buckets and was trained for that single job. Jev sorts almost any text into buckets you define on the spot, without training a model for each job, and gives every bucket a probability.

You can ask the models behind ChatGPT or Claude to do Jev's job today: give them the email and the three choices, and they will pick one. You get a written answer you have to trust or parse, at chat-model speed and price. Jev's pitch is doing that one job faster, cheaper, and with a probability you can set thresholds on, such as "send anything under 70% to a person." Whether that is worth switching for is most of the debate covered below.

How Jev works

TypeSafe calls Jev a "System One" model, after Daniel Kahneman's term for fast, intuitive thinking. In that framing a chat model is System Two, reasoning out loud one token at a time, and Jev answers without that step.

A call has two parts. The state is whatever your application knows, sent as a string, a JSON object, or an array: a support ticket, a search result, a lead record. The questions are typed. TypeSafe's documentation says the model reads the state once and evaluates every question against it in parallel, rather than generating an answer token by token. There are three question types.

Here is the Team question from the example above as the API sees it, trimmed from TypeSafe's API reference to one of its three questions. You send the state and the answer options:

POST https://api.typesafe.ai/v1/systemone
{
  "model": "jev-latest",
  "state": "Help! My payouts have been failing for 3 days.",
  "questions": {
    "department": {
      "type": "choice",
      "instructions": "Which team should handle this?",
      "criteria": {
        "billing": "Payments, invoicing, refunds",
        "technical": "Bugs, outages, integrations",
        "sales": "Pricing, upgrades, new accounts"
      }
    }
  }
}

You get back a pick, a probability for every option, and a confidence value:

{
  "model": "jev-1.13.0",
  "answers": {
    "department": {
      "type": "choice",
      "choice": "billing",
      "probabilities": { "billing": 0.88, "technical": 0.12, "sales": 0.0 },
      "confidence": 0.81
    }
  }
}

TypeSafe trained Jev with a method it calls Reinforcement Learning for Calibrated Decisions (RLCD). Calibration means that across many answers where Jev says 0.8, it should be right about 80% of the time. That is a claim about groups of predictions. It says nothing certain about any single answer.

Developers reach Jev through TypeSafe's own API, where the current version is jev-1.13.0, or through Vercel AI Gateway, Cloudflare, and OpenRouter. TypeSafe's documentation allows 64k tokens per request (32k of data plus the longest question), while Vercel and OpenRouter list 32K, so check the limit on the route you use.

The company spent about two years in stealth and raised a $40M seed round led by DCVC, SiliconANGLE reported. Forbes put the valuation at about $200M, citing a person familiar with the deal; TypeSafe has not published one. Founder and CEO Diogo Almeida worked at OpenAI on reinforcement learning from human feedback, InstructGPT, ChatGPT, and GPT-4. His co-founders are CTO Erik Gafni and COO Sasha Sheng.

Why the hype

Three things landed in the same week: a price close to zero, answers in under a second, and demos from people developers already follow.

Attention came first. Almeida's launch post on X passed 38 million views within a week, and TypeSafe's post ending the waitlist on September 20 drew another 1.4 million. Vercel reports Jev as the fastest-adopted model in AI Gateway history: nearly 13% of paid teams used it within 24 hours, about twice the share the GPT-5.6 family reached and more than six times Fable 5.1's. Those are Vercel's figures from its own gateway, and Vercel sells access to the model.

The demos that spread fastest:

• A safety reviewer. Vercel CEO Guillermo Rauch reported that Jev, tested as the safety reviewer in fx, was "up to 18x faster (p95) and more accurate" than the model doing that job.
• Context compaction. Alex Volkov ran Tamara Tran's compaction plugin and reported it cut his Claude session from nearly 1M tokens to 86K in about a second.
• A browser agent. Browser Use co-founder Gregor Zunic reported a flight search that took 7 seconds and cost $0.0039.
• Bulk labelling. Hassan El Mghari classified 1,018 AI research papers into 24 topics for $0.08 in total, at a median of 256 milliseconds per call.

Each is a single run by the person who built it, which shows what a good day looks like but is not a benchmark.

The speed and cost multiples are where the numbers drift. TypeSafe's press release says "up to 100 times" faster and less expensive. Its launch post reports 193.6x faster and about 445x cheaper on its internal workflow evaluations. LangChain's post says 200x faster inference and 400x lower cost. That is three sets of figures in one week, and all of them trace back to TypeSafe's own testing. TypeSafe also chose not to publish results on public benchmarks and tells users to build evaluations on their own data. For a decision model that is a reasonable position. It also means no outside party can reproduce the headline multiples yet.

The pushback was just as loud. Nathan Flurry's post calling Jev "just a really smart switch statement" became one of the most shared reactions of launch week. On Hacker News, when a commenter described Jev as "basically a zero-shot classifier," Almeida replied "exactly right!" The compaction demo drew the sharpest critique: Theo Browne called it "a terrible compaction strategy," arguing that Jev judges each step of a session without seeing what that step returned. The argument is about whether calibration and price make an old idea worth switching to.

Open-source copies arrived within days. openjev-sglang serves a Jev-compatible API from an open Qwen model, Bespoke Nimble publishes the data and recipe for an open version built on Qwen3.5-9B, and Convai's Laya pitches itself as a faster, multilingual alternative you can run yourself.

What Jev is for

Jev fits anywhere software makes the same bounded decision many times a day, and a slow or expensive answer costs more than an occasional wrong one.

• Routing. Send each request to the right model, queue, or agent tool.
• Support triage. Category, urgency, and whether a person needs to see it now.
• Search filtering. In a chatbot that looks things up before it answers (a setup called RAG), check whether each passage it found actually answers the question before the chatbot writes its reply.
• Checks on other models. Whether a citation supports its claim, whether an output breaks a policy, whether an extracted field matches its source.
• Agent loops. The small "continue, retry, or stop" decisions an agent makes dozens of times per task.

TypeSafe's own use-case map has sections for advertising and lead generation. The jobs we would test first:

• Search-term negatives. Read the search terms report and flag the queries to exclude, with a probability you can set a threshold on.
• Creative tagging. Label every ad by hook, format, offer, and claim, so performance can be grouped by what the ad says.
• Claim and brand-safety checks. Flag regulated claims in fintech, health, or crypto copy before it ships, for a person to approve.
• Ad-to-page match. Check whether the landing page delivers what the ad promised.
• Lead scoring and routing. Score inbound leads against an ideal-customer rubric and route them to the right rep.
• Content QA and internal links. Gate drafts on a checklist and suggest which pages should link to which.

How well any of these works depends more on how you write the questions than on the model. The phishing test below puts a number on it.

What is proven and what is still a claim

Three independent tests were published within days of launch. They are more useful than the launch material, and each comes with caveats its author wrote.

Close to a fine-tuned model on banking intents

GitHub user simonmesmith ran jev-1.13.0 on the 3,080 test messages of BANKING77, a public dataset of banking customer messages sorted into 77 intents. Jev scored 92.40% (2,846 of 3,080). A fine-tuned BERT model published by Casanueva et al. in 2020 scored 93.66%. Macro-F1 was 0.9235, the mean time per message was 0.533 seconds and the p95 was 0.713 seconds, both including retrieval and network time, and the experiment cost $0.7671 in API calls. The author lists the caveats up front. This was not zero-shot: each message came with 24 similar labelled examples retrieved by BM25 search, plus definitions for all 77 intents. BANKING77 is public, so it may sit in Jev's training data, and nobody has audited that. The BERT figure is a published number, not a rerun.

Ahead of zero-shot BERT models

A second project, by GitHub user zhuyansen, compared Jev with BERT-family zero-shot classifiers on seven evaluation sets. Jev won on all seven, by 0.05 to 0.13 accuracy on the standard benchmarks. On AG News it scored 0.865 against 0.763 for a clean DeBERTa zero-shot model. The most useful result is on 258 arXiv papers submitted after Jev's release, which cannot be in its training data: Jev led by 0.30 accuracy. On PAWS, a paraphrase set, the gain was 0.03 AUC. The caveats: against bge-m3 embeddings Jev was a statistical tie on Banking77 (0.712 vs. 0.722), bare label names cost it 0.13 on Banking77, and the post-release set is small and uneven across topics.

Question design decides the result

The most practical finding so far comes from XenoSpectrum, which ran Jev on 2,000 emails to detect phishing. Asked as a single question, "Is this email phishing?", Jev scored 62.6%. Split into five narrower questions and combined, it scored 89.4%. With the five answers weighted by a logistic regression fitted on half the emails, it scored 95.0%. Same model, same emails, and a 32.4-point spread that came entirely from how the questions were written. The 95.0% was not a statistically significant improvement on Haiku's 93.2% on the same task.

JevHaiku, for comparison

One question62.6%

Five questions, combined89.4%

Five questions, fitted weights95.0%

Haiku, same task93.2%

With Jev, writing the questions is most of the work of building the classifier. Budget time for question design and for a test set drawn from your own data, which is what TypeSafe tells users to do anyway.

Where Jev does not fit

Jev picks answers from options you define. It does not explain its choice, write text, or read images.

• No audit trail. There is no chain of thought, no "why" string, and no feature attribution. If a customer or regulator asks why a lead was rejected, Jev's output alone will not answer them.
• Text only. No images, audio, or video. Tagging ad creative means transcribing or describing it first.
• English first. TypeSafe says other languages work, with lower quality.
• Known weak spots. TypeSafe's notes on jev-1.13 list numbers, dates, and reading wording too literally.
• Calibration describes averages. A 0.9 is right about nine times in ten across many answers. Any single 0.9 can still be wrong.
• It is not a control. Jev does not replace authorization, deterministic validation, or human review. Use it to triage and filter, and keep a person or a hard rule on anything with legal, financial, or safety consequences.
• It does not generate. If the output is prose, code, or a reply, use a language model. Jev can sit in front of one and decide when it runs.

22 sources

Introducing System One Models & Jev typesafe.ai 1 Jev models, pricing, and limits docs.typesafe.ai 2 Example use cases docs.typesafe.ai 3 Jev 1.13 known weak spots docs.typesafe.ai 4 TypeSafe AI exits stealth with $40M siliconangle.com 5 Jev adoption on AI Gateway vercel.com 6 Building a harness with Jev langchain.com 7 Guillermo Rauch on Jev as a safety reviewer x.com 8 Alex Volkov on context compaction x.com 9 Gregor Zunic on a Browser Use flight search x.com 10 Hassan El Mghari on labelling 1,018 papers x.com 11 Nathan Flurry's "smart switch statement" post x.com 12 Almeida's reply on Hacker News news.ycombinator.com 13 Jev on BANKING77 github.com 14 Jev vs. zero-shot BERT models github.com 15 Question decomposition on 2,000 emails xenospectrum.com 16 Diogo Almeida's launch post x.com 17 TypeSafe opens Jev to everyone x.com 18 Theo Browne on the compaction demo x.com 19 openjev-sglang github.com 20 Bespoke Nimble, an open Jev x.com 21 Laya by Convai laya.convaiinnovations.com 22

Frequently asked questions

Related: AI consulting · How to get ChatGPT to recommend your business · Does llms.txt work?