What Is Jev?
Jev is the first System One Model released by TypeSafe AI, a San Francisco AI lab founded by Diogo Almeida — who previously worked on the instruction-following research at OpenAI that became ChatGPT — alongside Erik Gafni and Sasha Sheng. Launched into early access on September 15, 2026 alongside a $40 million seed round led by DCVC, Jev does not generate text. Instead, given a piece of unstructured input and a set of typed questions you define in advance, it returns typed answers with calibrated probability and confidence scores — output meant to be consumed directly by software, not read by a person.
How Does Jev Work?
Traditional large language models generate output autoregressively — one token at a time, sampling from a probability distribution over the entire vocabulary at every step. Jev skips that process entirely. It evaluates the full input in parallel and returns structured, type-safe decisions in a single pass, using a training method TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD). Because its output space is constrained to a schema you define upfront, Jev cannot return a value outside that schema — no invented fields, no malformed JSON, no fabricated tool names.
Jev exposes three query primitives:
- Choice — pick one option from a closed list (up to 255 options), with a probability distribution across all candidates.
- Score — place the input on an ordinal scale (e.g. severity 0–4), returning the expected level and the full distribution across ranks.
- Noul — a direct yes/no question, returning a calibrated probability between 0.0 and 1.0 rather than a blunt boolean.
How Fast and Cheap Is Jev?
TypeSafe reports 70–500ms end-to-end response times — 40 to 200x faster than frontier LLMs on comparable classification and routing tasks. Pricing is aggressive: $0.042 per million input tokens, with output tokens free, since Jev isn't generating text to bill by the token. By September 18, three days after launch, Vercel reported Jev as the fastest-adopted model in its AI Gateway's history.
What Is Jev Actually For?
Jev is not a chat model and isn't meant to replace one. It's built for the decisions that happen inside software: classifying support tickets, routing requests, scoring risk, extracting structured fields from unstructured text, and gating branching logic in an agent's execution path. TypeSafe also positions it as a guardrail layer — a fast, cheap way to verify or score an LLM's output before it reaches a user or triggers an action.
Vantaverse's Take: Where Jev Fits in Agent-Native Software
We build MCP servers and agent harnesses for a living, and Jev's pitch resonates with something we've argued for a while: not every decision inside an agent's execution loop needs a full LLM call. When an agent has to route a ticket, score a lead, or decide whether a tool call's output should be trusted, a fast, calibrated, type-safe decision model is a better fit than a conversational model that might hallucinate a field name. We expect models like Jev to sit alongside LLMs inside production agent harnesses — the LLM handles reasoning and language, and a System One model handles the split-second structured decisions the LLM is comparatively slow and expensive at. If you're already building agent-native software and hitting latency or cost walls on routing and classification steps, this is the layer worth evaluating first.
Is Jev Open Source?
No. Jev is a closed, hosted model — you call it through TypeSafe's API or SDKs, and its weights are not published. If open weights and self-hosting matter to you, the closest alternative is Laya, an open-source System One model from Convai Innovations, which we cover in a separate breakdown — including exactly where it beats Jev and where it doesn't.