SHELFCRITTER AI Digest

COMPANY

TypeSafe AI

companytopic-note

Overview

TypeSafe AI is the vendor behind Jev — a new “System One” / decision model class returning typed probabilistic outputs (yes/no, category, calibrated score) rather than free-form text. TypeSafe AI is the only vendor shipping this shape today, with Jev priced at $0.042/M input tokens with free output and running 40–200× faster than small frontier LLMs on inference. The category is one-vendor so far and, until a second vendor (Cohere, Anthropic, or open-weights via Kev) ships a comparable product, is more accurately read as a product than a category.

Timeline

  • 2026-09-22-AI-Digest — Simon Willison flagged Jev on Sept 21 — a new “System One” / decision model class from TypeSafe AI returning typed probabilistic outputs rather than free-form text, priced at $0.042/M input tokens with free output, 40–200× faster inference than small frontier LLMs. Jev’s non-autoregressive architecture returns calibrated confidence scores on typed slots — the differentiator Willison highlighted is that the confidence numbers are actually usable in downstream routing decisions, unlike LLM-produced probabilities that don’t calibrate. Jared Palmer’s open-source Qwen3.5 refactor Kev (0.8B / 4B / 9B checkpoints, no Jev outputs used in training) shipped within 24 hours.

Key Developments

  1. Jev — Non-Autoregressive Decision Models: TypeSafe AI’s Jev is architecturally distinct from “prompt an LLM for JSON output” or Cohere-style classify endpoints — non-autoregressive with calibrated confidence scores on typed slots, 40–200× faster inference as a direct consequence of the architecture rather than an optimization pass. Priced at $0.042/M input tokens with free output.

  2. One-Vendor Category, 24-Hour Open-Weights Compression: TypeSafe AI is the only vendor shipping decision models today. Kev‘s Qwen3.5 refactor arrived within 24 hours of Jev’s release with an explicit “no Jev outputs used in training” claim — if it holds under scrutiny, another data point that the closed→open replication window for a novel model shape is now measured in hours, not weeks.