Guide 46 · Basics
What Is Jev? TypeSafe's System One AI Model Explained
Jev doesn't chat — it decides. TypeSafe's new model class trades text generation for 70–500ms structured decisions your software can use directly.
By Abhit Tomar · Updated November 16, 2026 · 17 min read

Short answer
Jev is TypeSafe AI's first 'System One' model, launched in early access on September 15, 2026. Instead of generating text like ChatGPT or Claude, it takes unstructured state plus a fixed set of typed questions and returns structured decisions with calibrated probabilities — in 70–500ms per call, at $0.042 per million input tokens with outputs free.
Jev in one paragraph
TypeSafe AI was founded by Diogo Almeida — an ex-OpenAI researcher who worked on the instruction-following methods behind ChatGPT — and spent two years in stealth. On September 15, 2026, the company opened early access to Jev, the first public model of a new class it calls “System One” models: frontier models built to make fast, structured decisions that software can consume directly, rather than chat with humans.
The pitch, in TypeSafe's own words: think of Jev as a frontier-intelligence function call — unstructured state in, typed probabilistic decisions out. It spent launch day at the top of Hacker News, and within a week developers were building Doom bots, wiki-racing agents, and browser automation on it.
The audience is developers building software that decides, not people looking for a chatbot. Jev’s natural home is inside an application: a support system that routes a ticket to the right queue, a sales tool that scores a lead, a content pipeline that classifies what comes in, a guardrail that checks an agent’s planned action before it runs. In each case the questions are fixed in advance — the software knows the options — and what it needs from the model is a fast, honest answer about which option fits and how confident it is. That is a much narrower job than chatting, and Jev’s whole design is a bet that the narrow job is where most of the volume will be.
Decide, don't chat: how Jev differs from LLMs
Every number below comes from TypeSafe's launch post — this is the company's framing of its own product, not independent benchmarking:
- Outputs: LLMs generate strings — chat replies, code, and occasionally hallucinations that must be parsed and validated before software can use them. Jev returns type-safe structured values defined in advance, each accompanied by calibrated probabilities and confidence scores.
- Sampling: LLMs generate one token at a time, each conditioned on the last. Jev generates all outputs in a single parallel query.
- Speed: TypeSafe reports 70–500ms end-to-end per call for Jev — 40–200x faster than frontier LLMs on the same class of queries.
- Cost: LLM pricing ranges from $0.20 to $10 per million input tokens, with outputs roughly 5x pricier. Jev charges $0.042 per million input tokens — and output tokens are free, described as “too cheap to meter.”
- Confidence: LLMs tend to be overconfident and inconsistent when asked how sure they are. Jev communicates calibrated uncertainty with every output: higher confidence means higher accuracy, and similar inputs return similar answers.
- Training: LLMs are tuned with RLHF or RLVR (human preference / verifiable rewards). Jev uses a new method TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD), optimizing for epistemically honest probabilities.
Jev vs the competition: what it actually costs
Jev's headline pricing is $0.042 per million input tokens, with outputs free. Here's what that means next to the frontier chat models — worked out for 1M input + 500K output tokens in a month, from our price table (verified against each provider's official pricing page on October 3, 2026):
One honest caveat before the numbers: this is apples-to-oranges by design. Jev's outputs are tiny typed decisions, not essays or code — it can't do what the chat models do, so “cheaper” only matters for decision-shaped work.
- TypeSafe Jev: $0.042 — output tokens free
- Mistral Large 3: $1.25
- Gemini 3.8 Flash: $2.63 (introductory rate through Dec 31, 2026)
- DeepSeek V4 Pro: $3.30 (peak hours; half off-peak)
- Grok 4.7: $5.00
- Claude Sonnet 5.5: $7.00
- GPT-6.1 Sol: $7.00
Why 'can't hallucinate' — and the honest caveat
TypeSafe's boldest claim is that Jev “can't hallucinate.” The logic: hallucination is a property of string generation, and Jev never generates strings — its outputs are schema-constrained values, so type errors are, in the company's words, mathematically impossible.
Skeptics have a fair counter, and it's worth stating plainly: a model constrained to three categories can still confidently pick the wrong one. Not generating text eliminates hallucinated prose, not wrong decisions. TypeSafe itself publishes nuance notes alongside its demos — including that its eye-catching figures (193.6x faster, 444.6x cheaper) come from its own workflow evals, built by its own team, and likely sit at the higher end of real-world gains.
There is a second, quieter point hiding in the claim. Because Jev’s outputs are typed values with probabilities attached, the software around it can do something chatbots cannot: set a threshold. Below a certain confidence, the system can ask a human, fall back to a slower model, or simply refuse to act — and because the confidence is calibrated, that threshold means something predictable. An overconfident model makes thresholds useless; a calibrated one makes them a real control. That is the sense in which the honest-caveat version of the claim still matters: Jev does not eliminate wrong answers, but it gives the system around it a reliable signal for when to doubt them.
What's in a name
“System One” borrows from Daniel Kahneman's Thinking, Fast and Slow — the distinction between fast, intuitive System 1 thinking and slow, deliberate System 2 reasoning. TypeSafe's argument: most software doesn't need deliberation, it needs fast judgment with honest confidence.
“Jev” honors William Stanley Jevons, the economist behind the Jevons paradox — the observation that making a resource dramatically cheaper tends to increase total demand for it. The bet: every order-of-magnitude drop in the cost of intelligence unlocks orders of magnitude more use cases.
Proof it's real: what developers are building
The most striking early demo isn't from TypeSafe at all. Browser-use's open-source jev-ultrafast project (21,800+ GitHub stars, MIT-licensed) wires Jev into a browser agent: one Jev request picks both the operation (click, type, scroll…) and the target element per decision cycle, with a small conventional LLM used only to write text when typing is needed. The result: a real Google Flights search from Zürich to London completed in about 7 seconds.
TypeSafe's own demos include a Doom bot making decisions from structured game state roughly 10 times a second, and a Wikipedia-racing agent choosing between hundreds of links per step. Community reports also mention emerging open-weights alternatives in the same spirit — Laya, Nimble, and Kev — for developers who want to self-host rather than use the API.
What the demos prove, and what they do not, is worth separating. They prove the latency claim is real enough to change what a loop can do: a browser agent that decides in under half a second can afford a decision per step, and a game bot deciding ten times a second is simply not a shape a chat model can take. They do not prove production reliability — the demos are built by enthusiasts and by TypeSafe itself, the pricing is early-access pricing, and the independent benchmarking that would settle the speed and cost multiples has not caught up yet. Since launch, TypeSafe has shipped Jev 1.14 (October 14, 2026), and the category already has a second vendor: OpenAI announced its own Decisions API at DevDay on September 29, 2026. The honest read is that Jev is a real, working product with an unproven scale story.
Limitations to know
- Jev only works when the possible answers are defined up front. Open-ended questions, creative writing, and anything needing an explanation are outside its design — that missing capability is the product, not a bug.
- Access is early-access via waitlist; this is brand-new infrastructure, not a mature platform.
- The headline speed and cost multiples are TypeSafe's own measurements. Independent, large-scale benchmarking hasn't caught up yet.
- It's built for decisions inside software — routing tickets, scoring leads, classifying content, guardrailing agent outputs — not as a chatbot you'd talk to.
Why this matters if you use Muse
Muse is a conversational agent: brilliant at open-ended work, reasoning through ambiguity with you. Jev is the complement — the fast, cheap decision layer that can sit inside software and answer “which of these options, and how sure are you?” in milliseconds. The emerging agent pattern is exactly this split: let the decision model classify and score, and let ordinary code (plus a chat agent for the fuzzy parts) authorize and act.
As agents take on more real-world work — shopping, scheduling, triaging — expect this division of labor to become the standard architecture. Jev is the first high-profile attempt to productize the decision half.
Jev questions, answered
What is Jev? TypeSafe AI's first “System One” model (early access since September 15, 2026) that returns structured, typed decisions with calibrated probabilities instead of generating text.
Is Jev an LLM? Not in the usual sense. It's a new model class trained with RLCD for calibrated decisions, with parallel rather than sequential sampling — built for decisions, not dialogue.
How fast is Jev? TypeSafe reports 70–500ms end-to-end per call, claiming 40–200x the speed of frontier LLMs on decision-shaped queries.
How much does Jev cost? $0.042 per million input tokens, with output tokens free — per TypeSafe's published pricing.
Can Jev hallucinate? It can't hallucinate text, because it never generates text. It can still make wrong decisions within its fixed options — confidence scores are meant to tell you when to trust it.
How can I try Jev? Through the early-access waitlist at typesafe.ai, or via the open-source browser-use/jev-ultrafast project on GitHub, which needs a TypeSafe API key.
What is a System One model? TypeSafe's term for models optimized for fast, structured, software-consumable decisions — named for Kahneman's fast System 1 thinking.
Where to go next
Start with the official sources to try Jev — and if you build agents and automations, our open-source skills library is linked below too.
- TypeSafe — official website — join the Jev early-access waitlist.
- Introducing System One Models & Jev — TypeSafe blog — the launch post this guide is built on.
- browser-use/jev-ultrafast — GitHub — the open-source ultrafast browser agent (MIT license).
- TypeSafe docs — API reference and evals.
- AI model cost calculator — our tool — plug in your own token volumes and compare Jev against six frontier models side by side.
- AI model cost dashboard — our tool — interactive charts: per-token rate bars, a monthly bill explorer, and cost-vs-volume curves.
- awesome-muse-skills — our GitHub repo — 899 original open-source skills for AI agents and automation builders.
- Muse AI Personal Agent — how conversational agents handle approvals and actions.
What would you actually build with Jev?
The pattern to look for is decisions your software already makes badly, slowly, or expensively. A support desk that routes tickets by keyword could ask Jev which queue a ticket belongs to and how sure it is, escalating only the uncertain ones to a human. A sales team could score every incoming lead the moment it arrives, so the hottest ones get a call first. A content platform could classify uploads — spam, safe, needs review — at a cost so low the check can run on everything. An agent that acts in the world could run its planned action past Jev first: is this the right tool, the right account, the right amount — and is the model confident enough to proceed? In each case the value is not that Jev is smarter than a chat model; it is that the decision is cheap enough and fast enough to put in the loop, and honest enough about its confidence to know when to hand off.
Is Jev a replacement for ChatGPT or Claude?
No, and the confusion comes from calling both of them models. A chat model takes an open-ended question and reasons through it in text; Jev takes a fixed set of options and returns a typed decision with a probability. They answer different questions, and the interesting systems use both: the chat model handles the fuzzy parts — understanding what the user wants, drafting the message, reasoning through an edge case — and the decision model handles the structured parts, thousands of times a day, in milliseconds, for fractions of a cent. If you need a conversation, Jev is useless. If you need a thousand routing decisions a minute with honest confidence scores, a chat model is the wrong tool. The split is the point: Jev is not a smaller chatbot, it is a different organ in the same body.
Keep reading
01 · BasicsWhat Are ChatGPT Dots? The Complete Guide to OpenAI's Always-On AgentsOpenAI's Dots keep working after you close ChatGPT. What they are, what they can do, how to set one up, what they cost, and where the limits are — in one place.Read guide
03 · BasicsWhat Is Muse AI? A Clear Beginner's GuideUnderstand what Muse is, what it can do, and where it fits among personal AI agents.Read guide