TypeSafe Jev 1.13: a model for application decisions

A customer message needs to go to returns or logistics. The application wants a category, while a conventional chatbot may generate a paragraph explaining its choice. TypeSafe AI designed Jev around work where software needs a structured decision.

Jev 1.13 is available on OpenRouter as typesafe/jev-1.13. The listing gives a release date of 18 September 2026, a 32,000-token context and a price of USD 0.042 per million input tokens, with no output-token charge. These details were checked on 28 September 2026; we check the current rate before quoting a deployment.

Inputs and outputs

You supply a description of the situation and questions with allowed outcomes. The response contains typed answers and probabilities. TypeSafe presents Jev as giving up free-form text generation in favour of this interface.

An example application could ask which department owns a ticket, whether it needs urgent escalation and whether the request is complete. These are separate questions about the same case. The application uses their answers in workflow rules, such as sending an ambiguous ticket to an operator.

That is a proposed use case. It is not a Syntalith Jev benchmark or evidence of performance on a particular company’s queue.

Three question types and API access

choice selects from your categories, noul returns the probability of yes, and score places an item on an ordered scale. For a ticket, these might identify the department, whether a refund was requested and the urgency level. The model accepts text, including a JSON object, but does not directly process images or recordings. It also does not explain its decision.

According to the OpenRouter documentation, an OpenRouter key is enough; a separate TypeSafe account is unnecessary. Requests use the Decisions API at POST /api/alpha/decisions or the TypeSafe SDK-compatible POST /api/v1/systemone. Keep the key on your application server. Independent questions about the same state can share one request; they cannot read each other's answers.

In a pilot, we record the model version returned by the API. A latest-version alias can change results, so thresholds need another check on labelled data after a version change.

Why the speed claims attract attention

Many applications do not need a long answer. They need to recognize intent, check a condition or choose the next step quickly. Jev addresses this demand through a different approach to structured outputs.

TypeSafe publishes its own speed and cost comparisons. The authors describe the tasks, conditions and limitations. We treat those as vendor measurements that justify a trial. We do not transfer their percentage gains into promised savings for a customer.

An application also spends time on networking, input processing, validation and other services. Faster model computation does not necessarily reduce the entire workflow by the same proportion.

Valid format and correct decisions

Separate two kinds of error. A model can return an invalid structure that software cannot read. It can also return a permitted category that is wrong for the case.

TypeSafe’s type-safety claims concern output structure. “Logistics” can be a valid data value and still be the wrong destination for a billing complaint. A high confidence score also needs evaluation on company examples. The implementation therefore includes assessment, escalation thresholds and handling for missing information.

Can Jev run locally?

OpenRouter offers a hosted TypeSafe service. The materials we checked do not provide downloadable Jev 1.13 weights for self-hosting. For a private processing requirement, we check provider terms, location and availability. The Jev name does not establish offline operation.

Separate local projects draw inspiration from this approach, including Kev and TinyJev. They have their own weights, training recipes and limitations. Our local alternatives article explains the distinction.

How Syntalith can help

We select a decision from the workflow that can be scored clearly. We compare rules, a small model and a general-purpose baseline, adding Jev where access and terms fit the requirements. We measure accuracy, escalation rate and cost per accepted decision.

If you have a ticket or document classification queue, describe it to Syntalith. We can prepare an evaluation, choose a hosting approach and adapt a local model to your company’s categories.