TypeSafe AI releases Jev model — and it's not an LLM

Startup TypeSafe AI emerged from stealth on September 15 with the Jev model — a 'System One Model' that doesn't generate text but instead processes state data and typed queries in parallel, outputting typed answers with confidence scores.

Compiled by Sergey KostenchukPublished 2026-09-18Updated 2026-09-19

2026-09-18 Research OpenAI

🤖 TypeSafe AI releases Jev model — and it's not an LLM

The startup introduced the 'System One Model': instead of generating text, it processes state data and typed queries in parallel, outputting answers with confidence scores. The creator is Diogo Almeida, a former OpenAI employee and co-inventor of RLHF.

🌍 Attack on 'one model for everything': in agentic pipelines, most LLM costs and latency are routine tasks (routing, classification). The company claims 193.6x speedup and 444.6x savings compared to frontier LLMs, but benchmarks are compared against other frontier models, not a standard.

👤 If you're building agents, Jev is a candidate for moving routing, classification, and data extraction out of LLMs at ~$0.0004 per decision. Every confirmed ~25x faster and 580x cheaper than Claude for data extraction. Try it: typesafe.ai.

Source 1: https://forkast.news/typesafe-ais-jev-is-not-an-llm-and-that-may-be-the-point/ Source 2: https://www.every.to/chain-of-thought/typesafe-ai-jev

Sources