TypeSafe Launches Jev for Fast, Structured AI Decisions

TypeSafe AI has launched Jev, a model built for software to make structured choices such as classification, routing and scoring. Rather than generate a string of text token by token, Jev returns typed probabilistic decisions in parallel—a design TypeSafe says can make AI cheap and fast enough for production decision paths where a chatbot would be unnecessary overhead.

Jev is available in early access after TypeSafe spent two years in stealth, according to founder Diogo Almeida. Almeida says the company built a new architecture, parallel sampler and training method called Reinforcement Learning for Calibrated Decisions, or RLCD, around the goal of giving software a decision it can use directly.

A constrained answer instead of an open-ended one

The contrast with a conventional large language model is deliberate. Language models generate flexible text, which can serve as chat, code or a structured response, but software must parse and validate that output before using it. Jev’s possible outputs and their structure are defined in advance, so the model cannot produce free-form text or stray outside the required schema.

That limitation is central, not a footnote. Jev is intended for what TypeSafe calls “smart if-statements”: decisions inside workflows where a brittle hand-written rule might otherwise classify an input, select a route, assign a score or trigger a branch. It is not presented as a replacement for an assistant that needs to write an explanation, compose code or carry on a conversation.

The speed pitch rests on a specialized workflow

TypeSafe lists Jev at $0.042 per million input tokens and says output tokens are free. It reports end-to-end response times from 70 to 500 milliseconds, and says its workflow evaluations produced results as much as 193.6 times faster and 444.6 times cheaper than the comparison systems used in those tests.

Those comparisons are TypeSafe’s own results, and the company supplies unusually direct qualifications. Its workflow authors came from the model-capabilities team, creating possible bias. The reference answers were averages from GPT-6 Astra and Fable 5.1, while the competing language models were constrained through TypeSafe’s own structured-output wrapper, which the company says is accurate but slower and more expensive than requesting decisions without probabilities.

Confidence is part of the product claim

RLCD is meant to make Jev’s confidence scores correspond to its accuracy: higher confidence should mean a more accurate answer, TypeSafe says. The company argues that this calibration, paired with schema-constrained outputs, makes the model better suited to automation than a system that returns an answer without reliably indicating uncertainty.

Where TypeSafe says Jev fits

• Classifying, routing, scoring, extracting or branching on unstructured inputs within ordinary software workflows.
• Processing large datasets into features and insights through repeated structured decisions.
• Real-time applications where TypeSafe says roughly 100-millisecond responses can affect user experience.

The unresolved test is whether developers can turn that narrowness into dependable systems. TypeSafe says Jev can avoid type errors because its schemas constrain what it returns, but it is asking early users to identify both the decisions they need automated and the places where the model falls short. Its launch makes a specific bet: many valuable AI tasks may be decisions inside software, not conversations with it.