decision model layer ai agents stack sprawl a neapolitan ice cream brick with its middle layer slid out

As AI Agents Grow, Vendors Carve Out Decision-Making as a Separate Model Layer

October 9, 2026
Artificial Intelligence
Comment off

AWS, Cloudflare and OpenAI have joined TypeSafe in shipping decision models, small models that make the bounded choices inside AI agents while a larger model does the reasoning. InfoWorld’s sources warn the new layer brings stack sprawl, confidence scores that do not transfer between vendors, and decision schemas that encode business policy. We compare Jev, Clef, Strands Decider 2B, GLiNER2.5-Decide and the Decisions API on price and latency, then set out the calibration, governance and cost-per-decision work that decides whether the savings hold.

A ChatGPT Inventor’s New Kind of AI Model Is Thrilling Developers

September 21, 2026
Artificial Intelligence
Comment off

Jev reached nearly 13% of Vercel’s paid teams within twenty-four hours, more than double any previous model launch on the AI Gateway, and TypeSafe AI briefly ran out of API capacity. Vercel, Bryo AI and Earendil have now said publicly what they measured. An independent phishing benchmark scores the same model at 62.6% or 95.0% depending only on how the question is framed, and the out-of-distribution calibration numbers decide whether any of it survives contact with your data.

ChatGPT Pioneer Launches Jev Model for Programmatic Logic

September 16, 2026
AI News
Comment off

TypeSafe AI has left stealth with $40 million led by DCVC and Jev, a model that does not chat, write code or generate a single word of prose. Founded by InstructGPT co-author Diogo Almeida, it returns typed decisions with calibrated probabilities in 70 to 500 milliseconds, at $0.042 per million input tokens with output billed at nothing. The speed and cost claims survive outside testing; the accuracy question does not get a clean answer.