Jev Only Picks From the Answers You Give It

On September 15, TypeSafe AI released Jev along with a $40 million round led by DCVC. The company was founded by Diogo Almeida, who co-invented RLHF and InstructGPT at OpenAI. I nearly scrolled past the announcement because I assumed it was another chatbot with a bigger benchmark number.

Jev is what they call a decision model. You give it a piece of state (an email, a log line, a JSON blob) plus a batch of typed questions. Each question comes with an enumerated answer set, which can be yes/no, multiple choice or a rubric score. In one API call you get back a typed answer for every question, each with a calibrated probability and confidence. It generates no tokens at all. A new architecture and a parallel sampler read the input once and score every question in the same forward pass instead of threading them through one autoregressive stream. TypeSafe frames this as "System 1" judgment, fast and intuitive, against the "System 2" token-by-token grind of a normal LLM.

What the constraint buys you

Speed is nice but it's not what caught my attention. The only legal outputs are the options you defined, so the model has no way to invent a tool that doesn't exist or return a category you never listed. Ask it a four-option question and it can't hallucinate a fifth answer. If you've ever written defensive parsing for a malformed tool-call response because the model decided to freelance, you know why I care about this.

You also get a real probability back instead of a vibe. That lets you set thresholds and send low-confidence cases to a human or a bigger model.

The pricing follows from the same design. TypeSafe charges $0.042 per million input tokens with output free, which makes sense when there's barely any output to pay for. They also claim sub-200ms latency on multimodal inputs. Those are vendor numbers and I'd benchmark them on my own traffic before believing them.

Where I'd use it

Jev is not a chatbot replacement and it would be a mistake to treat it as one. It can't summarize or draft anything. It can't explain itself or reason out loud either. What it handles is the boring high-volume step that a lot of systems currently burn a full LLM call on. Which agent should handle this ticket? Is this message a policy violation? Should this tool call be allowed to run? Which of these six intents is the user expressing? Anywhere your software repeatedly makes a fuzzy decision from a finite set of actions, a cheap calibrated classifier in front of the expensive generative model is sensible architecture. In an agent system I'd use Jev as the router and the gatekeeper around the model doing the thinking, not as the brain itself.