TypeSafe AI is an AI lab officially positioned as "building machine-native intelligence infrastructure," with the goal of enabling decision-making and automation within software. Its first public model is Jev in the System One Model series, currently offered in early access.
[Purpose and Positioning] TypeSafe AI believes that reinforcement learning from human feedback (RLHF) has led LLMs to be optimized to cater to human preferences, thereby achieving superhuman performance in instruction following, which is what is now called "chat." However, RLHF brings inherent problems such as mode dropping, overconfidence, and lack of reliability, which often make LLMs require humans-in-the-loop. Therefore, TypeSafe AI chooses the opposite research direction, building a class of models used natively by machines—System One Models—and adopting a new architecture, a new sampler, and a new training algorithm: Reinforcement Learning for Calibrated Decisions (RLCD).
[Main Capabilities] According to the official introduction, Jev's core characteristic is outputting "Decisions, not strings": returning typed outputs/typed decisions that software can act on. Every decision comes with the model's own estimate of confidence (calibrated confidence). Officially, it is reliable, fast, and type-safe, and emphasizes "Zero Hallucinations"—because every decision carries a confidence estimate, software can execute when confidence is high and escalate when confidence is low. Users can set thresholds in code to decide when to act autonomously and when to request review, and compose these decisions into larger workflows while controlling how intelligence is used.
[Performance and Comparison] Data displayed on the official website claims that its solution is 193.6 times faster and 444.6 times cheaper than comparison solutions (based on a System One task workflow, with a proof link attached). The page uses "Completed in 0.114s" versus "Completed in 8.566s" as an example, and states that ""Jev's intelligence per dollar is off the charts," with the chart caption "Workflow Intelligence vs. Cost." The official source also states that its input price is lower than Claude Fable 5.1 (per billion input tokens).
[How to use] The officially stated usage is: send structured questions to Jev, obtain decisions with type, probability, and confidence, and have software act accordingly. Users can control the boundary between autonomous execution and human review by setting thresholds, and combine these decisions in code to build larger workflows. The official website provides "Try our first System One Model, Jev, in early access," as well as a blog, FAQ, and contact information for getting started or asking questions.
[Applicable scenarios] Built for automation, it is suitable for workflow scenarios that require calibratable decisions inside software, have requirements for reliability and speed, and want to use confidence to manage uncertainty. The official source says its "Workflow Intelligence vs. Cost" performance stands out.
[Limitations and not yet provided] The official materials do not provide specific pricing figures, complete API documentation, a list of available platforms, or detailed usage steps. Its page is marked Version 0.01, dated October 12, 2026 (including wording such as Monday, Oct 12, 2026), and contains some encoded content (such as a C language code snippet, a Base64 text, and a base64 image); this content is insufficient to confirm as feature descriptions. Jev is currently in early access and may not yet be fully open.
[FAQ (officially listed)] The official FAQ headings include: What are System One Models? What is Jev? Is Jev a smaller LLM? How is it different from JSON mode or structured outputs? Why is Jev so fast and cheap? Can Jev be made faster? Are these prices temporary or subsidized? What is Jev good at, and where does it struggle? Does Jev still make mistakes? Is Jev deterministic? How do I get started or ask questions? However, the materials only list the question headings and do not provide the corresponding answer content.