TypeSafe AI launches Jev, a typed‑decision model promising zero hallucination and massive cost savings

TypeSafe AI announced the early‑access release of Jev on 15 September 2026 after two years of quiet development. The company positions Jev as the flagship of its newly defined System One Models, a class of AI systems built to make structured decisions inside software applications rather than to generate conversational output.

What distinguishes System One Models

System One Models differ from traditional large language models by requiring a typed request that includes the current state and a predefined set of possible choices. The model processes this input and returns a decision together with calibrated probability scores, allowing downstream code to apply automation thresholds or trigger human escalation when confidence is low.

According to TypeSafe, the architecture relies on a parallel sampler and a training regime called Reinforcement Learning for Calibrated Decisions (RLCD). The company claims this combination yields decisions that stay within the prescribed type space, which it markets as "zero hallucination" because the model cannot produce output outside the defined schema.

Performance claims and pricing

TypeSafe reports that Jev delivers 193.6 × faster inference and 444.6 × lower cost on System One workflows compared with its own baseline models. These figures stem from internal demonstrations and have not been independently verified.

The pricing model is straightforward: $42 for every one billion input tokens, with no charge for output tokens at launch. This contrasts with typical per‑token pricing that bills both input and output, potentially lowering total cost for decision‑heavy workloads.

Limitations and operational considerations

Despite the zero‑hallucination claim, TypeSafe cautions that the selected decision may still be incorrect. The model can only choose among the supplied options, so errors arise from mis‑calibration or ambiguous inputs rather than from generating unrelated text.

The number of parameters, detailed architecture, training data, and a full model card were not disclosed at launch. Prospective users must therefore conduct their own validation, especially to verify that probability estimates are well calibrated for their specific domain.

• Validate calibration on representative internal datasets
• Define clear automation thresholds based on probability scores
• Implement a fallback human review workflow for low‑confidence cases
• Monitor decision outcomes for drift and retrain as needed

Jev is designed for classification and decision tasks where the set of possible outcomes is known in advance, such as fraud detection, loan approval, or routing tickets. It is not intended for free‑form content creation, summarisation, or any use case that requires open‑ended generation.

Organizations planning to adopt Jev should allocate resources to test the model on their own data, tune the RLCD training pipeline if needed, and establish governance policies that define when a machine decision can be acted upon without human oversight.

In practical terms, an English‑speaking enterprise can replace a portion of its rule‑based decision engine with Jev, achieving faster response times and lower compute spend while retaining a safety net through probability‑based thresholds. The shift promises more adaptable decision logic that can evolve with data, but it also introduces new responsibilities around calibration monitoring and human‑in‑the‑loop design.

Developers integrating Jev will need to adapt their APIs to send typed payloads that enumerate the current context and the exhaustive list of permissible actions. The response payload will contain a single chosen action identifier and a numeric confidence value that can be directly consumed by automation scripts.

Early adopters are encouraged to document any divergence between the model’s confidence scores and actual outcomes, as this feedback loop is essential for refining the RLCD process and ensuring that the promised cost efficiencies materialise in production environments.

Overall, Jev represents TypeSafe AI’s attempt to carve out a niche focused on reliable, typed decision making rather than the broader generative ambitions of most large language models, with the explicit goal of reducing hallucination risk and operational expense.