TypeSafe AI Debuts Jev, a Model That Plays Doom for Machines

Summary

TypeSafe AI, a startup with $40 million in funding, has released Jev, a model designed to interact with software and machines rather than people. Instead of generating natural-language responses, Jev returns typed, structured decisions with probabilities, allowing developers to avoid parsing and validating free-form text. A decision might assign a customer-service request to billing, technical support, or sales, while the model can also control a Doom game from structured game-state data. Jev is described as a System One model built on TypeSafe AI’s Reinforcement Learning for Calibrated Decisions architecture. The company says its parallel-processing design produces responses in 70 to 500 milliseconds, compared with sequential token generation in conventional LLMs; a company demo showed 0.114 seconds for Jev versus 8.566 seconds for OpenAI’s GPT-5.6 Terra. TypeSafe lists Jev at $0.042 per million input tokens with no output charge, while the cited GPT-5.6 Terra pricing is $2 per million input tokens and $12 per million output tokens. The company says Jev is suited to automation, real-time decisions, large-scale classification, input verification, and model harnesses. Its “hallucination-free” claim is limited: structured output may prevent fabricated prose and malformed tool calls, but it can still make incorrect decisions. The article also notes that greater reliability in deeply nested agent workflows could make the model useful where latency guarantees matter, while warning that the same capabilities could eventually be applied to sensitive actions such as drone targeting.