TypeSafe AI Launches Jev: A Machine-First Model Built for Speed and Structured …

TypeSafe AI on Tuesday released Jev, a new AI model designed to communicate with software rather than people. The startup, which has raised $40 million in funding, argues that traditional language models are poorly suited for machine-to-machine workflows and built Jev to fill that gap.

Unlike ChatGPT or Claude, which return natural language text, Jev outputs typed probabilistic decisions. A query routing a customer service ticket might return “billing: 0.08, technical: 0.85, sales: 0.07” along with a confidence score. For a human, that format is awkward. For a software system, it eliminates the parsing and validation required to process unstructured text.

RLCD Architecture and System One

The model relies on a different architecture called Reinforcement Learning for Calibrated Decisions, or RLCD, which powers what TypeSafe calls its System One models. Diogo Almeida, co-founder and CEO of TypeSafe AI, helped invent reinforcement learning from human feedback while at OpenAI and worked on ChatGPT. He founded TypeSafe to pursue machine-native AI rather than models built for human conversation.

“I spent years working on models designed to make AI better at interacting with people,” Almeida said in a statement. “But if AI is going to fundamentally change how work gets done, people can’t be the only consumers of intelligence.”

Speed and Performance

The architecture processes queries differently than standard large language models. Models like GPT-5.6 Terra predict the next token sequentially. System One returns all outputs at once through parallel processing, which TypeSafe claims produces response times between 70 and 500 milliseconds. In a demo on TypeSafe’s website, Jev responded in 0.114 seconds. GPT-5.6 Terra took 8.566 seconds for the same query.

Pricing Structure

The pricing structure reflects the difference in scope. Jev costs $0.042 per million tokens for input and nothing for output. GPT-5.6 Terra charges $2.00 per million tokens for input and $12 per million tokens for output. Compared to Fable 5.1, TypeSafe estimates Jev is roughly 238 times less expensive.

TypeSafe claims Jev is hallucination-free. That framing requires context: Jev does not generate natural language, so it cannot produce fabricated quotes or legal citations in the way a chatbot might. However, it can still return incorrect structured responses. The company argues that a hallucinated tool call in an AI agent is inconvenient, but becomes critical in systems with latency guarantees or deep dependency chains.

Use Cases

Use cases TypeSafe envisions include AI automation software, real-time applications requiring quick decisions, AI map-reduce jobs for classifying large datasets, verification of AI inputs, and AI model harnesses.

The Name: William Stanley Jevons

The name Jev honors William Stanley Jevons, a 19th-century English economist known for the paradox that improved efficiency in resource usage tends to increase rather than decrease consumption. TypeSafe is betting that as tokens become cheaper and more efficient, overall token consumption will grow—assuming the market for AI interactions expands as broadly as the market for energy. That assumption remains unsettled, given that many people with access to AI tools have not found compelling uses for them or avoid them over ethical concerns.

Jev can play Doom when fed structured data describing game state. The demonstration suggests the model’s ability to handle real-time decision-making with structured inputs. Whether that capability extends to higher-stakes applications like autonomous targeting systems remains an open question the company has not addressed.