AI tools · September 20, 2026

TypeSafe AI's Jev: the 'System One' decision model that's not a chatbot

A San Francisco startup called TypeSafe AI just came out of two years in stealth with a $40 million seed round led by DCVC and a new model called Jev — and it's deliberately not a chatbot. Launched September 15, 2026, Jev is the first so-called 'System One Model': instead of generating text, it takes unstructured state plus typed questions and returns typed, probabilistic decisions — Choice, Score, and Null primitives that your software branches on directly.

The difference is speed and cost. TypeSafe says Jev answers in 70 to 500 milliseconds end to end, using a new model architecture plus a parallel sampler that outputs all decisions in a single pass — no autoregressive token-by-token generation. It's priced at $0.042 per million input tokens, with output free because it's 'too cheap to meter.' Under the hood is a training method TypeSafe calls RLCD — Reinforcement Learning for Calibrated Decisions — which optimizes for calibrated probabilities on structured decisions rather than the RLHF and verifiable rewards used by GPT-5.6 and Opus 5-class models, as DataCamp's independent explainer lays out. Developers hit it at POST https://api.typesafe.ai/v1/systemone.

The idea is named after Kahneman's System 1 fast thinking and economist William Stanley Jevons of Jevons-paradox fame — and the implication is that cheap, fast machine judgment makes every agent workflow economically viable. Two honest caveats: TypeSafe's own blog admits it 'can't prove' the pricing isn't subsidized, and every performance figure is vendor self-reported with no independent reproduction yet. The company's founder is CEO Diogo Almeida, formerly of OpenAI, with Erik Gafni as CTO and Sasha Sheng as COO.

Why it matters

If agents are the future of software, they need a cheap judgment layer — every decision an agent makes today costs an expensive, slow LLM call. A model that returns calibrated decisions in milliseconds for fractions of a cent could make always-on agentic workflows practical for small businesses and builders, not just frontier labs. Watch for independent benchmarks before treating the speed claims as gospel.

Key facts

TypeSafe AI (San Francisco, founded 2024) emerged from stealth on September 15, 2026, with a $40M seed round led by DCVC
Jev is the first 'System One Model' — it gives up text generation entirely and returns typed, probabilistic decisions (Choice / Score / Null primitives) that code branches on directly
End-to-end response time: 70–500 ms, per the launch blog's spec table (vs seconds to minutes for frontier LLMs)
Pricing: $0.042 per million input tokens; output tokens are free ('too cheap to meter')
Trained with RLCD (Reinforcement Learning for Calibrated Decisions), optimizing calibrated probabilities instead of RLHF-style reward — DataCamp independently explains the category and method
Developer endpoint: POST https://api.typesafe.ai/v1/systemone; founded by CEO Diogo Almeida (ex-OpenAI) with CTO Erik Gafni and COO Sasha Sheng
Caveat: all performance claims are vendor self-reported with no independent reproduction yet; named after Kahneman's System 1 and economist William Stanley Jevons

Sources

Spot an error? We correct quickly and note it. Contact us via the contact page.

← More AI news