TypeSafe AI launches Jev, a typed classifier for faster agent routing

TypeSafe AI, a San Francisco lab that came out of stealth on September 15, 2026, has launched Jev, its first public model. Jev is not a large language model but a typed classifier built to replace prompt-and-parse routing calls inside AI agent systems. The company raised $40 million in seed funding led by DCVC.

What Jev does differently

Where a large language model decodes text one token at a time, Jev processes a block of program state together with a set of typed questions in a single forward pass, returning a label, a probability for each possible option, and a confidence score. TypeSafe calls this a "System One" model to separate it from LLM-style reasoning. Developers define up to 255 possible options, send the state to be classified, and get back a structured answer rather than free text that needs parsing.

TypeSafe says the model is calibrated through a training process it calls RLCD, so that reported probabilities track how often the model is actually correct — a property the company describes, with qualification, as avoiding open-ended hallucination, since results are constrained to the defined enumeration rather than free-form text.

Speed and cost claims

According to TypeSafe AI, Jev responds in roughly 70 to 500 milliseconds, which the company says is 40 to 200 times faster than routing the same decision through an LLM, at a company-stated cost of about $0.001 per decision. Vercel's engineering team has said it measured an 18-times speed-up after switching its production safety classifiers to Jev — a third-party data point, though still a single vendor account rather than an independent benchmark.

Early framework support

Support for Jev has appeared quickly across developer tooling:

• LangChain exposes it through a dedicated TypeSafeClassifier component.
• Pydantic AI treats it as a classifier that needs the same measurement discipline as any other production classifier.
• Neuron AI, a PHP agent framework, built a general classifier interface with TypeSafe as its first implementation.

What to watch

Jev is a classifier, not a general-purpose reasoning engine, and it inherits classifier-style risks: adversarial or unusual input text can shift its output, so teams are advised to pair it with deterministic checks rather than rely on it alone for security-relevant routing decisions. For MSPs and platform teams evaluating agent architectures, the more durable takeaway is the pattern of separating classification from generation, rather than any single vendor's speed or cost numbers, which remain self-reported and not yet independently benchmarked.

Sources: TypeSafe AI — Jev classifier, Eastern Herald.