TypeSafe Jev System One Decision Model
Released by TypeSafe AI on 15 September 2026 and described as Jev, the first public System 1 decision model — a new class of frontier models that returns calibrated probabilities and type-safe structured values instead of generating text; developers define questions in natural language with primitives like choice, score and null and get machine-readable probabilistic outputs back in 70-500ms, priced at $42 per billion input tokens ($0.042 per million) with output tokens free, trained with what the team calls RLCD (reinforcement learning for calibrated decisions) and available through early access by waitlist
Tool Overview
Features, steps and FAQ below
Features
- ✓ Structured decisions instead of text generation: TypeSafe says Jev is the first public System 1 model, where developers define questions in natural language with primitives like choice, score and null and the model returns type-safe structured values rather than strings, slotting into ordinary software as probabilistic decision rules
- ✓ Calibrated probabilities and confidence: TypeSafe says every output comes with calibrated probabilities and confidence scores, where higher confidence means higher accuracy and similar inputs return similar answers; because the output structure is defined in advance the model makes no type errors and, per TypeSafe, cannot hallucinate
- ✓ Faster and cheaper: TypeSafe says end-to-end response time is 70-500ms, with input tokens priced at $42 per billion ($0.042 per million) and output tokens free; in its published workflow evals it reports speedups of about 193.6x and cost savings of about 444.6x versus reference models
- ✓ Parallel sampling and a new training method: TypeSafe says Jev uses a parallel sampler that generates all outputs in a single query, making it highly efficient and hardware-aware, and is trained with RLCD (reinforcement learning for calibrated decisions) rather than conventional RLHF or verifiable-reward RL
- ✓ Use cases and status: TypeSafe says Jev fits AI-powered workflows and smart if-statements (classify, route, score, extract or branch), map-reducing over big data, 100ms real-time applications, and verification, scoring, judging, guardrails and jailbreak detection; it is available today in early access, with the team bringing developers off the waitlist as quickly as possible
How to Use
- Join the waitlist on the TypeSafe website to get early access
- Define your question in natural language, describing the possible outputs and structure with primitives like choice, score and null
- Send unstructured state such as text and get type-safe structured outputs back with calibrated probabilities and confidence scores
- Embed the outputs as composable fuzzy decision rules for classification, routing, scoring or verification, and integrate with the system as documented
FAQ
What is Jev?
TypeSafe AI describes Jev as the first public System 1 decision model, a new class of frontier models that returns calibrated probabilities and type-safe structured values instead of generating text, giving developers machine-readable probabilistic outputs in 70-500ms.
How is it different from ordinary LLMs?
TypeSafe says existing LLMs optimize for human preference or verifiable rewards and output strings, while Jev optimizes for calibrated decisions, returns type-safe structured values with calibrated probabilities, samples in parallel rather than token by token, and makes no type errors.
How fast and cheap is it?
TypeSafe says end-to-end response time is 70-500ms, with input tokens at $0.042 per million and output tokens free; in its published workflow evals it reports about 193.6x faster and 444.6x cheaper than reference models.
What is it good for?
TypeSafe says it suits AI-powered workflows and smart if-statements (classify, route, score, extract or branch), map-reducing over big data, 100ms real-time applications, and scoring, judging, verifying, guardrailing and detecting jailbreaks of LLM prompts, reasoning traces or outputs.
How can I use it today?
TypeSafe says Jev is available in early access today, so developers can join the waitlist on its website and the team is bringing developers off it as quickly as possible; it also offers open-source tooling like the System One LLM adapter and documentation.