The Pulse
TypeSafe launches Jev for fast structured AI decisions
TypeSafe AI has released Jev, an early-access model designed to return structured, probabilistic decisions rather than generated text. The company says Jev delivers 70–500 millisecond response times and can cost up to 444.6 times less than
AI.info Team ·
“We built a new stack entirely focused on automation: with a new model architecture, parallel sampler for maximum efficiency, and training method we call Reinforcement Learning for Calibrated Decisions (RLCD).”
Diogo Almeida, founder, TypeSafe
TypeSafe AI released Jev on September 14, introducing what it calls a “System One” model designed to make fast, structured decisions that software can consume directly. The company is opening the model to developers through an early-access program after two years in stealth.
Jev is aimed at tasks such as classification, routing, scoring, extraction and branching inside software workflows. Rather than producing a string that an application must parse and validate, it returns type-safe values defined in advance, along with probabilities and confidence scores.
Jev replaces generated text with typed decisions
TypeSafe describes Jev as a “frontier-intelligence function call”: unstructured state goes in, and typed probabilistic decisions come out. The model does not generate ordinary text, which the company presents as a deliberate trade-off aimed at reducing latency, cost and the risk of malformed output.
Existing language models typically generate one token at a time, with each token conditioned on the previous one. Jev instead samples its outputs in parallel, a design TypeSafe says is better suited to software that needs many independent decisions rather than a conversational response.
The company also says Jev cannot produce type errors because the model’s possible outputs and structure are defined in advance. TypeSafe distinguishes that guarantee from a broader claim about correctness: a result can match its required schema while still making a bad decision, which is why the system returns confidence information with every answer.
TypeSafe reports 70–500 millisecond response times
TypeSafe reports end-to-end response times of 70 to 500 milliseconds for Jev. In its comparison with frontier language models, the company says those figures can make Jev 40 to 200 times faster on “System One shaped queries,” although the comparison depends on the task, hardware and workflow.
The company lists Jev’s input price at $0.042 per million tokens and says output is free. TypeSafe also claims that its workflow evaluations show Jev reaching the same general level of intelligence as larger models while occupying a different cost-and-speed trade-off.
TypeSafe’s homepage attributes headline gains of 193.6 times faster and 444.6 times cheaper to the workflow evaluations published alongside the launch. Those figures come from the company’s own test design, not from an independent benchmark, and TypeSafe says the workflows were created by members of its model capabilities team.
The evaluation uses reference models, not a fixed answer key
TypeSafe’s workflow evaluation treats a software workflow as the object being tested. Each model receives the same workflow, and Jev’s outputs are compared with the average probabilities produced by two larger reference models identified by the company as GPT-6 Astra and Fable 5.1.
The approach is intended to measure how models behave inside code rather than how well they match a single ground-truth label. TypeSafe says the evaluation rewards fine-grained probabilities and decomposed decisions, which are then combined into discrete branches by the surrounding workflow.
That setup gives the company a useful way to test its product’s intended use, but it also limits what the results establish. TypeSafe acknowledges that its team built the workflows and that the language models were constrained through a wrapper designed to produce structured decisions compatible with Jev’s API.
Early access puts the model in developers’ hands
Jev is available now to developers admitted through TypeSafe’s early-access program. The company is asking users to test which decisions they can automate, where the model performs well and where it falls short.
TypeSafe’s launch examples include a Doom-playing system that reacts to structured game state and a Wikiracing agent that chooses among potentially hundreds or thousands of links at each step. The company says Jev supports choices with cardinality up to 255; larger selections use a two-stage scoring and choice process.
The product’s immediate target is not a replacement for general-purpose chat. TypeSafe is positioning Jev as a software component for workflows where applications need fast, constrained outputs, explicit uncertainty and predictable data structures. Its first public release therefore makes a narrower promise than a conventional language model: fewer words, more decisions, and a response format code can use without parsing prose.