Jev is a “System One” model. You give it a piece of text (the state) and a few typed questions. It answers every question at once with a typed value and a probability, never with prose. Your code takes it from there.
An LLM writes its answer one token at a time, and your code parses the text afterwards. Jev scores all three questions in one parallel pass. The clocks use the timings from TypeSafe's own homepage demo (0.114 s vs 8.566 s), played in real time.
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.” Costs are TypeSafe's demo figures ($0.000081 vs $0.013880). Vendor numbers, not an independent benchmark.
Pick a ticket or write your own, choose which questions to ask, and send them. Watch each question run in its own lane, at the same time. The answers here are simulated in your browser with a small keyword model so you can see the shape of a real response. This page does not call TypeSafe's API.
Every Jev question is one of three types. They can be mixed freely in a single call, and each one is judged on its own against the same state, so adding questions barely changes latency and one question can't contaminate another.
A statement about the state. Jev returns one number between 0 and 1.
A list of named options with descriptions. Jev returns the winner, a probability for every option, and a confidence. Up to 255 options.
An ordered rubric. Jev returns a probability per level, a numeric score (can land between levels), and a confidence.
Because the output is constrained to your schema, Jev can't return a malformed type or an option you didn't list. It can still pick the wrong option. That is what the probabilities are for.
Jev is trained with Reinforcement Learning for Calibrated Decisions (RLCD). The goal: when it says 0.9, it should be right about 9 times in 10. That lets your code automate the confident decisions and route the rest to review. Below are 200 simulated decisions. Drag the threshold, then switch to an overconfident model to see why calibration matters.
TypeSafe's advice is to keep each question atomic, the kind of call an expert makes in a few seconds, and to build the hard judgment in code. Here a security alert runs through two Jev calls. Your if statements decide what happens; thresholds are numbers you can change, not prompt wording. Values are illustrative.
| Maker | TypeSafe AI, San Francisco. Founded 2024 by Diogo Almeida (ex-OpenAI, InstructGPT co-author). Out of stealth Sept 15, 2026 with a ~$40M seed led by DCVC. |
|---|---|
| Model class | “System One” model, named for Kahneman's fast System 1. “Jev” nods to the Jevons paradox: cheaper decisions get used far more. |
| Training | RLCD, Reinforcement Learning for Calibrated Decisions, with its own architecture and sampler. TypeSafe says training used synthetic data only. |
| Interface | POST /v1/systemone with state + questions; Python SDK typesafe-sdk; also on Cloudflare Workers AI as typesafe/jev. |
| Claimed latency | 70–500 ms end to end |
| Launch price | $0.042 per million input tokens; output free |
| Limits | 32,000-token context; Choice up to 255 options (bigger spaces need two stages); no written rationale for any answer. |
| Access | Early-access waitlist. |