TypeSafe AI Releases Jev: A System One Model
TypeSafe AI launches Jev, a System One AI model available via hosted API in early access, providing calibrated decisions and typed outputs at 42 dollars per billion input tokens.
Stock photo for illustration only, not from the actual event
- TypeSafe AI releases Jev, a System One model via API in early access
- Jev utilizes a new stack including RLCD for calibrated confidence answers
- Internal evaluations claim 193.6x faster speed and 444.6x lower cost
TypeSafe AI has officially announced the release of Jev, a new artificial intelligence system model built as a System One architecture. The model is currently available as a hosted API behind a waitlist in early access. The company has not yet published model weights, parameter counts, or any self-hosting options.
The name draws inspiration from Daniel Kahneman's distinction between fast intuition and slow reasoning. The TypeSafe team argues that models tuned via RLHF were optimized for human preference, which generated conversational chat, overconfidence, and mode dropping. These flaws necessitate keeping a human in the loop for critical tasks.
Under the hood, Jev runs on a novel technology stack comprising a new architecture, a parallel sampler, and Reinforcement Learning for Calibrated Decisions (RLCD). Despite detailing these components, TypeSafe has not disclosed the specific architecture powering the system.
Stock photo for illustration only, not from the actual event
Interactions with the model occur through a single endpoint at POST https://api.typesafe.ai/v1/systemone. The request body carries the state, model identifier, and a map of questions. According to the documentation, the system supports three question types, executing them in parallel and in isolation against the identical state. TypeSafe notes that adding more questions barely impacts response time, and a Choice question supports up to 255 options.
Every Choice and Score answer includes a confidence value ranging from 0 to 1, derived by TypeSafe from the shape of the probability distribution. In a documentation example, billing wins with a confidence of 0.84, while the overall confidence sits at 0.596 because technical still holds 0.159. The documentation outlines three operational paths: act automatically on high confidence, review moderate levels, and route low confidence to a human. Thresholds should scale appropriately with the cost of an incorrect action.
"Zero hallucinations means schema matching is guaranteed. The 0% figure is not empirical. Answers can still be wrong."
TypeSafe AI
The strategic shift toward typed, calibrated decisions rather than conversational text addresses a fundamental limitation of traditional Large Language Models. By outputting structured data alongside concrete confidence scores, Jev enables downstream automated systems to execute conditional logic natively without requiring brittle natural language parsing layers. This architectural approach is particularly valuable for enterprise and industrial workflows where strict data schema compliance and risk quantification are critical.
Priced at 42 dollars per billion input tokens, Jev compares against existing LLMs which typically range from 0.20 to 10 dollars per 1 million input tokens. In recorded demonstrations, Jev completed execution in 0.114 seconds at a cost of 0.000081 dollars, whereas GPT-5.6 Terra required 8.566 seconds at 0.013880 dollars. These figures derive from TypeSafe's workflow evaluations. The team clarifies that zero hallucinations signifies guaranteed schema matching, though the 0 percentage is not empirical and answers can still be factually incorrect.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment