Decision model JEV hits 1T tokens daily, quarter of Fortune 500 on board

大模型之路 · wechat · 2026-10-08

TypeSafe's decision model JEV processes 1 trillion tokens daily three weeks after launch, with a quarter of Fortune 500 companies integrated. It generates no text—only probability-calibrated structured judgments (Noul/Choice/Score primitives)—trained via RLCD so stated confidence is statistically trustworthy, enabling automation pipelines with confidence thresholds.

Four core use cases: ticket triage (Clef hits 94.2 macro-F1 on BANKING77), risk gating, browser automation (APUS open-sources fast-browser-use), content moderation and model routing (millisecond complexity routing to differently-priced models).

Cost: per 1,000 classifications JEV costs $0.0248 vs GPT-6 Luna $0.0921 and Claude Opus $2.88; per million messages $19-44 vs $344-1025 for Gemini/Haiku. Median latency 194ms, Cohen's κ agreement with three major LLMs is 0.80. Requests above 0.9 confidence (2/3) exceed 99% agreement with LLMs—safe to auto-execute, a "fast/slow division of labor."

Security caveat: adversarial inputs can sway JEV's verdicts (an rm -rf interception test was manipulated), so safety decisions still need independent sandboxing/permission gateways. The category is crowding: OpenAI Decisions API, Databricks aidecide, Cloudflare Clef, Amazon Strands Decider 2B. Limits: text-only, 64K token cap, unreliable at arithmetic, no general reasoning.

Related event: TypeSafe's Fast Decision Model Jev Goes Viral as OpenAI Chases(7 posts)→

Original post →

More from coding & agent

coding & agent channel →