Every production system built on a language model shares the same seam: the model hands back a string, and something downstream has to parse it, validate it, and decide what to do when the parse fails. TypeSafe AI's answer, published September 15, is to delete the string. Its first model, Jev, returns typed values drawn from a schema declared in advance, each carrying a calibrated probability.
Key takeaways
- TypeSafe emerged from stealth with roughly $40 million in seed funding led by DCVC, founded by former OpenAI researcher Diogo Almeida with Erik Gafni and Sasha Sheng.
- Jev prices input at $0.042 per million tokens and charges nothing for output, with end-to-end latency the company puts at 70 to 500 milliseconds.
- Because the output schema is fixed before the query runs, the company says type errors are structurally impossible rather than merely rare.
What a System One model is
TypeSafe calls the category System One: models built to make fast decisions that software consumes directly, rather than prose a person reads. Three pieces changed to get there - a different architecture, a parallel sampler, and a training method the lab names Reinforcement Learning for Calibrated Decisions.
The contrast with RLHF is the point. Where human-feedback training optimizes for responses raters prefer, and verifiable-reward training optimizes for outputs a checker can confirm, RLCD optimizes for honest probabilities attached to decisions. Almeida co-authored the InstructGPT work that underpins ChatGPT, so the departure is deliberate.
The sampling difference matters as much as the objective. Conventional models generate one token at a time, each conditioned on the last. Jev emits every field of its structured output in a single parallel pass, which is where the latency and cost claims originate.
The numbers, and what they rest on
The headline figures on the company's site - 193.6 times faster, 444.6 times cheaper - come from a purpose-built evaluation rather than a public leaderboard. TypeSafe fixes a workflow in code, runs every model through the identical compute graph, and scores each against reference probabilities taken from the average of GPT-6 Astra and Fable 5.1.
To the lab's credit, the caveats are published alongside. The four workflows were written by its own model capabilities team. Using frontier competitors as the reference biases results toward those competitors, which the company argues understates its own margin. And the zero type-error rate is not measured but guaranteed by schema matching - a claim, as the post notes, that a single counterexample would demolish.
Where it is supposed to fit
The pitch is not a chatbot replacement. TypeSafe positions Jev for the fuzzy decisions littered through ordinary software - classify, route, score, extract, branch - the places where hand-written logic turns brittle and a full large language model call is too slow or too expensive to sit in the hot path.
Its other suggested uses are map-reducing over large datasets, real-time interfaces where 100-millisecond budgets rule out a frontier call, and grading the output of other models as a judge or guardrail. One internal demo runs ten queries per second at roughly $7 an hour.
Why the framing lands now
The argument that models are too unpredictable to build on has been getting sharper, and not only from labs selling an alternative - unannounced weight changes behind stable API model names have made the same point from the other direction. A model that cannot produce a hallucination because it cannot produce free text is a blunt response to that complaint.
The cost of the trade is real, though. Anything requiring generated language - explanations, code, a written answer - is outside what Jev does at all, and TypeSafe concedes its published evaluations were run from its own laptops on the West Coast. Jev is waitlisted in early access, and the interesting number will be how the pricing holds once someone else is paying for the GPUs.
FAQ
What does Jev return instead of text?
Typed values selected from a schema the developer defines before the query runs, each accompanied by a calibrated probability and a confidence score. Because the possible outputs are fixed in advance, the response needs no parsing or validation step, and the model cannot return a field the schema does not allow.
How much does Jev cost?
Input tokens are priced at $0.042 per million, and output tokens are free - TypeSafe describes them as too cheap to meter. The company expects the price to fall rather than rise, though it acknowledges it cannot yet prove the rate is unsubsidized.
Is Jev generally available?
No. It is in early access for selected developers, with a waitlist at typesafe.ai. TypeSafe has published four workflow evaluations, including full queries and disagreements, for anyone wanting to inspect the methodology behind its speed and cost claims.
