Jev is a decision model built by TypeSafe AI, a San Francisco start-up. It does not write text: you send it a piece of state, such as a support ticket, an agent trace or a JSON object, together with a set of typed questions, and it returns structured answers with probabilities that your code can act on directly. As of October 2026 the current version is Jev 1.13, served through TypeSafe’s API and several gateways at $0.042 per million input tokens, with output tokens free.
This article draws on TypeSafe’s documentation, partner docs from LangChain, Vercel, Cloudflare and Pydantic, and press coverage. We have not run Jev in a client project, so every performance figure below is attributed to whoever measured it.
Who made Jev?
TypeSafe AI describes itself as an AI lab “building machine-native intelligence infrastructure for automation”. Its team page lists three founders: Diogo Almeida (CEO), Sasha Sheng (COO, formerly a research engineer at Meta/FAIR) and Erik Gafni (CTO, a repeat founder of Ravel and an early employee at Invitae and Freenome). The company says Almeida “co-invented RLHF and InstructGPT” and worked at OpenAI and Google Brain; TechCrunch reports that he left OpenAI two years ago to start TypeSafe.
TypeSafe came out of stealth on 15 September 2026 with about $40 million in seed funding led by DCVC. The launch press release gives 2024 as the founding year and names no other investors. Valuation figures circulating online are second-hand and unconfirmed by the company.
Is Jev an LLM?
TypeSafe says no. In its docs Jev is “TypeSafe’s flagship model and the first System One model”, and a System One model “evaluates a state and returns typed answers and probabilities” (System One concept page). The same page is explicit about what it leaves out: these models “do not write replies, produce code, or generate explanations of their reasoning.”
The term comes from Daniel Kahneman’s Thinking, Fast and Slow, where System 1 is fast and intuitive and System 2 slow and deliberate. The name Jev is a nod to the economist William Stanley Jevons and the Jevons paradox.
TechCrunch calls Jev “a new transformer-based model” that is “not a large language model”. TypeSafe has published neither the architecture nor a parameter count, and outside guesses that it sits on an open-weight LLM base are unconfirmed. On training, the docs name a method called Reinforcement Learning for Calibrated Decisions, or RLCD, and Almeida told TechCrunch the model is trained “exclusively on synthetic data”.
How does Jev work?
One request carries two things: a state (a string, a JSON object or an array) and a map of named, typed questions. Every question is evaluated “in parallel and in isolation against the same state”, and TypeSafe says adding questions “barely changes the response time”. Input is text only; images, audio and video are not supported.
This is the Python example from the official quickstart, quoted as published:
from typesafe_sdk import Choice, Noul, Score, TypeSafeClient
client = TypeSafeClient()
ticket = "Hi, I've been trying to connect my Stripe account for 3 days and the integration keeps failing. I'm losing sales. Please help ASAP."
response = client.system_one(
state=ticket,
questions={
"department": Choice(
instructions="Which team should handle this",
criteria={
"billing": "Payment or subscription issues",
"technical": "Bugs or integration problems",
"sales": "Pricing or account questions",
},
),
"frustration": Score(
instructions="How frustrated the customer appears",
criteria=[
"Calm, just stating facts",
"Frustrated but civil",
"Very angry, strong language",
],
),
"is_urgent": Noul(
instructions="The message conveys urgency or time-sensitivity",
),
},
)
print(response.answers["department"].choice) # "technical"
print(response.answers["frustration"].score) # 1.0
print(response.answers["is_urgent"].noul) # 1.0In the sample response on the same page, department comes back as technical with probabilities of 0.85 for technical and 0.15 for billing, and a confidence of 0.78. The whole call used 392 input tokens. There is nothing to parse, and the probabilities tell your code how sure the model is.
What are Jev’s question types?
There are three, and they can be mixed in one call (API reference).
Choice picks one option from a set you define and returns the chosen option, the full probability distribution and a confidence value. One Choice can hold up to 255 options. Larger sets need a slower two-stage process.
Score rates the state along an ordered rubric of at least two and at most 10 levels. It returns a score that is probability-weighted and can land between levels, a legend, the probabilities and a confidence. The docs warn against reading exact magnitudes out of a score.
Noul is the yes/no type. It returns the probability that the answer is yes in a field called noul, between 0 and 1, and accepts optional criteria.true and criteria.false descriptions. Noul has no confidence field. TypeSafe has not explained the name, and Vercel’s gateway API calls the same type boolean.
The confidence value on Choice and Score comes from the shape of the distribution: 1 when all probability sits on one outcome, 0 when it is spread evenly. TypeSafe adds a caveat for anyone setting thresholds: “Calibration is measured across groups of predictions; it does not guarantee that an individual answer is correct.”
Does Jev hallucinate?
The homepage says “Zero Hallucinations”. The launch post is more careful: the 0% figure for type errors and hallucinations rests on schema guarantees and is “not an empirical measurement”. In practice this means Jev cannot return an option you did not define or a malformed answer. It can still pick the wrong option. Armin Ronacher, CTO of Earendil, put it this way to TechCrunch: “At the end of the day, it delegates the hallucination problem a little bit to the user.”
How fast is Jev?
It depends on who is measuring. TypeSafe’s launch post gives 70 to 500 ms end to end, measured from team laptops on the US West Coast, and “40x to 200x faster than frontier models” on System One-shaped queries. The press release says “less than 100 milliseconds”, and the homepage claims 193.6x faster and 444.6x cheaper than LLMs on its own workflows.
Independent tests land lower on the multiplier. AY Automate measured a median of 0.33 s against 1.17 s for GPT-5.6 Terra, about 3.6x. A study summarised by Beri.net measured 239 ms from France against 687 ms for Claude Haiku 4.5. Cloudflare, which sells a competing model, measured a median of 524.1 ms and a p95 of 536.0 ms. A fair reading: medians of a few hundred milliseconds, and roughly 2 to 6 times faster than fast LLMs in outside tests, not 200 times.
How much does Jev cost?
TypeSafe charges $0.042 per million input tokens, which is $42 per billion, and does not bill output tokens (models page). The API still reports usage.output_tokens; it is just not charged. Cloudflare and Vercel list the same input price. Enterprise plans go through sales.
At list price the 392-token quickstart call costs about $0.000016, so a dollar covers around 60,000 calls of that size. TypeSafe also concedes in its launch post that it “can’t prove its pricing isn’t subsidized” and expects prices to fall.
How do you get access to Jev?
Jev launched with a waitlist, and demand briefly overwhelmed the API, according to TechCrunch. Around 20 September TypeSafe posted on X that “Jev is now available to everyone. No waitlist.” The current quickstart describes the path: create an account at console.typesafe.ai, take an API key from the dashboard, and try questions in the Playground.
Some operational details from the models page:
- Rate limits are listed as 100K tokens per second and 80 requests per second, with a warning that they “can change without notice”. Higher limits come with enterprise plans.
- The model ID
jev-1.13.0is current, and the aliasesjev-latestandjev-previewpoint to it. The docs advise pinning a version once you have tuned thresholds. - Jev is hosted only. The same weights serve every account, with no fine-tuning on customer data and no open weights or self-hosted edition.
- “Jev is not trained on customer requests or responses”, and zero data retention is available for enterprise customers.
- The context limit is given as 64k tokens per request, with 32k for the state plus the longest question. Cloudflare and Vercel list 32,000.
Which SDKs and integrations exist?
TypeSafe publishes a Python SDK (typesafe-sdk, Python 3.10 or later), a JavaScript/TypeScript SDK (@typesafe-ai/sdk, Node 20 or later), a plain HTTP API at POST https://api.typesafe.ai/v1/systemone, and an agent skill for Claude Code and other coding agents. The SDKs retry 429 and 529 errors with exponential backoff by default.
Partner integrations include a langchain-typesafe package in alpha with experimental routing middleware, typesafe-ai/jev on Vercel AI Gateway, typesafe/jev on Cloudflare Workers AI, TypeSafeModel in Pydantic AI, and listings on DigitalOcean and OpenRouter. How these fit into agent tool selection and model routing is covered in our guide to Jev agent routing.
Where does Jev fail?
TypeSafe documents this itself on its Jev 1.13 jaggedness, last reviewed on 2 October 2026. Its summary: “It may struggle with tasks that require additional levels of indirection. It can be quite literal in its understanding. It struggles with tasks that require numeric precision.” Listed failure modes include maths (“Jev is not a calculator”), counting, date and time comparison, long states full of irrelevant detail (“Jev suffers from context rot”), contradictory instructions, a lean towards whichever Choice option comes first, and adversarial content. On the last point the docs say plainly that state is data and Jev “does not treat it as hostile by default”.
That last one has been tested. VentureBeat reported an Octomind test in which a fake “pre-approved” field in the state moved Jev’s probability of blocking rm -rf ~/.ssh from 0.76 to 0.48. English is also the main training language; other languages “are handled but not equally well”.
On accuracy, TypeSafe’s own workflow evaluations put Jev at a mean of 67.8%, level with Claude Sonnet 5 and below GPT-5.6 Sol at 74.1%, with the reference labels taken from the consensus of two frontier LLMs, not from human annotators. We go through these numbers and the independent tests in Jev vs LLMs.
FAQ
What does “noul” mean in Jev?
Noul is TypeSafe’s name for its yes/no question type, which returns the probability that the answer is yes. TypeSafe has not published an explanation of the word. Vercel’s gateway API exposes the same type as boolean.
Is Jev open source?
No. TypeSafe serves Jev only as a hosted model through its API and partner gateways, and there are no open weights or on-premises edition. Open-source alternatives exist: Cloudflare released Clef under Apache 2.0 and says it is “fully Jev-API compatible”, and AWS released Strands Decider 2B.
Is Jev free to use?
Not entirely. Output tokens are free, but input costs $0.042 per million tokens, and the same rate applies on Cloudflare and Vercel. Enterprise plans with higher rate limits are arranged through TypeSafe’s sales team.
Can Jev write answers or summaries?
No. The docs say it is “not trained to generate text”. Jev returns choices, scores and probabilities; if your workflow needs a reply, a summary or code, an LLM still writes it.
Who are the founders of TypeSafe AI?
Diogo Almeida (CEO), Sasha Sheng (COO) and Erik Gafni (CTO), according to the company’s team page. TypeSafe was founded in 2024 and is based in San Francisco.
Where Jev fits in a real system
Jev is a narrow tool: cheap, fast enough for inline decisions, and openly weak on maths and indirection. Whether it pays off depends on how many decisions your system makes and how well they split into atomic questions you can test against labelled data. AI Superior designs and builds LLM applications and AI agents for B2B clients, including the evaluation and cost-control work that decides which model handles which step. If you are planning that kind of architecture, see our LLM development and consulting services.