Jev is the first model from TypeSafe AI, a San Francisco lab that came out of stealth on September 15, 2026 with $40M in seed funding. It is the most talked-about AI launch of the month, and it does something strange: it cannot write a single sentence.
Most frontier labs compete on better chat. TypeSafe removed chat entirely. Jev does not generate text, code or explanations. You send it data and a list of typed questions. It sends back one answer per question: a yes/no probability, one option from a list you defined, or a position on a scale you defined, each with probabilities. TypeSafe’s slogan sums it up: “Decisions, not strings.” Think of it as a smart if-statement. Ordinary code branches on values it can compute. It falls apart when the condition is a judgment, e.g., is this customer angry, is this claim risky? That gap is where Jev sits.
Key Features
- Typed Answers, Not Text: Jev supports three question types. Boolean estimates the probability something is true. Choice selects one option from a list you define. Score grades against an ordered rubric. Your code receives an actual boolean or enum, not a string to parse and validate. TypeSafe’s “zero hallucinations” claim means the schema match is guaranteed, not that the model is never wrong, Jev can still choose the incorrect class. Keep that distinction in mind.
- Calibrated Confidence: Every answer arrives with a probability. This is the feature that matters most for automation. In insurance underwriting, for example, Jev could review evidence about a property and estimate the likelihood it will catch fire. A workflow continues when confidence is high, while an ambiguous case routes to an underwriter. The confidence number is your escalation threshold, written in code.
- Parallel Evaluation: Regular language models generate text one token at a time, which the application then parses and validates. Jev evaluates all declared questions in parallel and returns typed answers plus probabilities directly.
- Speed and Price: TypeSafe says most calls complete in about 100 milliseconds. Input tokens cost $0.042 per million, output tokens are free. TypeSafe reports Jev was up to 193.6x faster and 444.6x cheaper than LLMs on its workflow evaluations.
- A New Training Method: TypeSafe built a new stack from scratch: new architecture, new sampler and a training method it calls Reinforcement Learning for Calibrated Decisions (RLCD). It named the class System One Models, after Kahneman’s fast System 1 thinking. Where RLHF optimises for human preference ratings, RLCD is designed to produce epistemically honest probability estimates.
- Already in the Gateways: Beyond TypeSafe’s own API, Jev reached Vercel AI Gateway within 48 hours of launch, plus Cloudflare, and OpenRouter in beta. AI SDK 7 exposes it through the experimental evaluate API. If you already route through a gateway, trying it costs one line of config.
Company Background
TypeSafe AI was founded in 2024 in San Francisco and spent about two years in stealth. It has three co-founders: CEO Diogo Almeida, a former OpenAI researcher who worked on RLHF, InstructGPT, ChatGPT and GPT-4; CTO Erik Gafni; and COO Sasha Sheng. Gafni previously worked at genomics startups Ravel, Invitae and Freenome, and Sheng is a former Meta and FAIR research engineer.
The founding story explains the product. “I spent years working on models designed to make AI better at interacting with people,” Almeida said. “But if AI is going to fundamentally change how work gets done, people can’t be the only consumers of intelligence.” The man who helped teach models to talk to humans now argues that most automation does not need talk at all.
The company announced a $40 million seed round led by deep-tech firm DCVC, and Forbes reported a post-money valuation of about $200 million. The launch landed hard: it sat at the top of Hacker News for most of launch day, and the founder’s announcement post passed four million views.
User Experience
- Getting Access: Jev is in early access. TypeSafe is admitting developers from a waitlist and has published a manifesto, public API documentation with a playground, a workflow evaluation site, and demos. The fastest route for JavaScript developers is the AI SDK: install ai@latest and call typesafe-ai/jev through the evaluate API.
- The Developer Model: You define the state (a string, object or array), then a map of named questions. The result object gives you typed answers directly, e.g., result.answers.refunded is a real boolean. No prompt engineering theatre, no JSON repair code, no regex.
- The Real Skill Is Question Design: Your results are only as good as your criteria. A vague Choice option produces a confident wrong answer. The work shifts from writing prompts to writing precise decision definitions, closer to designing a form than chatting with a model.
- Cost in Practice: One developer counted roughly 5,000 requests for about $2 on day one. Experimentation is effectively free.
- Known Friction Points: The benchmark caveat comes first. There are no independent evals at launch, and the headline numbers are TypeSafe’s own. Every’s independent test came out 25x faster, not 200x, and Jev caught 6 of 7 planted defects where Claude Fable 5.1 caught 7. Still impressive but a more honest baseline than the marketing. Second, scope: Jev cannot reason through multi-step problems, explain its answer, or handle anything requiring generated output. Complex workflows often require open-ended reasoning as an intermediate step, and that still needs an LLM. Third, TypeSafe itself describes Jev as “still in its early days,” and its own evaluation caveats are worth reading before making architectural bets.
Cost
Pricing is unusually simple. There are no tiers, no seats, no subscription.
Early Access (Pay As You Go)
- Jev via TypeSafe API: waitlist admission at typesafe.ai.
- Price: $0.04 per million input tokens; output tokens free.
- Includes: Boolean, Choice and Score evaluations with calibrated probabilities, API docs and playground.
- Via Gateways: Vercel AI Gateway, Cloudflare, OpenRouter (beta). Same model, billed through your existing gateway account.
Free output changes the economics of decision-heavy workloads. A classification call that costs a fraction of a cent on an LLM costs a rounding error on Jev.
In summary, Jev is not a better chatbot. It is a bet that most automation does not need generated text, it needs fast, cheap, calibrated judgment that plain code can consume. Classification, routing, triage, guardrails, escalation. For that slice of work, a purpose-built decision model beats a chat model with a JSON instruction bolted on.
It is a poor choice for anything requiring reasoning chains, explanation or generated content. And until independent benchmarks arrive, the vendor numbers deserve healthy scepticism.
But here is the reason to watch it regardless: the confidence score is a governance primitive. Today, most agentic systems make decisions inside opaque text generation, and human oversight is a policy statement rather than a mechanism. A typed decision with a calibrated probability gives you something you can actually govern, e.g., auto-approve above 0.90, human review below, and an auditable decision trail in between. The escalation threshold stops being a paragraph in a policy document and becomes a line of code. That is the shape AI governance needs to take, and Jev is an early, concrete example of it.