On September 15, 2026, TypeSafe AI unveiled Jev, a model designed for structured decision-making rather than text generation.
On September 15, 2026, TypeSafe AI introduced a new category of AI model known as System One Models. The first model in this category, named Jev, is engineered to revolutionize how software makes structured decisions. Unlike conventional AI systems which focus on generating text, Jev streamlines the decision-making process.
Most AI systems today are built around a generation interface. Users provide instructions, and the model generates tokens, which are then interpreted. This methodology, while flexible, introduces complexities when integrated into software that expects deterministic outputs.
Defining the System One Model
TypeSafe differentiates its System One Models with a clear operational framework:
- State: Data input from the application.
- Questions: Bound inquiries that the model must evaluate.
- Model Execution: The System One Model processes the questions.
- Typed Decisions + Probabilities: Outputs that can be directly consumed by the application.
- Application Logic: The final step where decisions are utilized.
This approach allows applications to provide structured data, such as customer requests or support inquiries, and receive concise, machine-readable decisions in return.
Jev’s Capabilities
Jev’s architecture enforces a significant shift in how AI interacts with software. Rather than generating text, it focuses on returning choices, scores, and probabilities. This allows for greater efficiency when making decisions about customer interactions, fraud detection, and risk assessment.
TypeSafe claims that Jev’s output can include probability distributions, for example:
billing 0.91
technical 0.05
account 0.03
other 0.01
This means developers can program responses based on confidence levels, automating actions or escalating issues as needed.
Efficiency Over Generation
Traditional language models generate outputs sequentially, which can be inefficient for tasks that do not require natural language processing. System One Models, including Jev, eliminate the need for sequential text generation, purportedly leading to considerable efficiency gains.
TypeSafe highlights that many AI workflows currently utilize generative capabilities for straightforward decision-making tasks, including:
- Is this message spam?
- Which department should receive this ticket?
- Should this tool call be approved?
- Is this document relevant?
- Does this response satisfy a policy?
By bypassing unnecessary prose generation, Jev can potentially enhance performance, offering faster response times.
Performance Metrics
TypeSafe positions Jev at a competitive price point of $0.042 per million input tokens. The system reportedly achieves response times between 70 and 500 milliseconds based on the complexity of the request. Internal evaluations suggest Jev can be up to 193.6 times faster and 444.6 times cheaper than traditional systems under specific conditions.
It is essential to interpret these claims critically. TypeSafe acknowledges the tests were conducted internally and may not reflect broader real-world performance.
Addressing Hallucination Claims
TypeSafe makes a notable assertion that Jev cannot hallucinate outputs. This means that it cannot generate arbitrary responses outside its defined parameter set. However, it is crucial to distinguish between structural failures and decision failures. While it cannot return invalid categories, it may still produce incorrect decisions within valid options.
This new paradigm proposed by TypeSafe AI is aimed at creating a model that enhances decision-making processes in software applications, steering away from generative outputs towards a more structured approach.