TypeSafe AI’s Jev answers with choices, never sentences

TypeSafe AI began moving developers off the waitlist for Jev on September 15. Jev is a model that never writes free text: the developer declares the possible answers and their types in advance, and Jev returns one of them with a confidence score that the application can use as is. A Testingcatalog rundown covers the pricing, the benchmarks and the caveats TypeSafe attaches to both.

Jev comes out of two years in stealth. Diogo Almeida, a former OpenAI researcher who worked on reinforcement learning from human feedback and early ChatGPT research, founded TypeSafe in 2024 with Erik Gafni and Sasha Sheng. DCVC led the $40 million seed round. Almeida’s launch post on X opened with the question behind the company: “After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?” His answer is a new way to train models, which he calls RLCD, and “a new type of frontier AI model”.

TypeSafe calls Jev its first System One Model, a fast decision layer for software. It takes unstructured context, text or structured program state, and the developer defines the available choices and the output structure up front. A chat model writes its answer one token at a time. Jev evaluates the candidate decisions in parallel and returns probabilities over them. The company says it can process hundreds of outputs in parallel from a single prompt and that its confidence scores are calibrated, so developers can decide when software acts autonomously and when it defers.

Jev handles up to 255 choices directly. Larger choice sets go through two stages: one scores each candidate independently, and the other selects among them. TypeSafe lists three components: a new model architecture, a parallel sampler and Reinforcement Learning for Calibrated Decisions, the method behind the RLCD acronym. Almeida says his earlier work at OpenAI covered instruction following and the conversational model methods that helped lead to ChatGPT.

TypeSafe says schema matching is guaranteed and type errors are mathematically impossible. It describes the problem it is solving as models that hallucinate, switch methods from one request to the next and add variability to systems that need predictable behavior. The claim that Jev cannot hallucinate applies only to the constrained output space, though. It does not mean every factual decision will be correct, and Jev can still pick the wrong option from a valid list.

Jev costs $0.042 per million input tokens, or $42 per billion, and output tokens are free. TypeSafe reports end-to-end response times of 70 to 500 milliseconds and says Jev can be 40 to 200 times faster on similarly intelligent System One-shaped queries. Other launch figures differ. Press materials carried by Techstartups and Yahoo Finance said “up to 100 times faster and less expensive” with sub-100 ms latency, while Almeida’s X post said “20-200x faster” and “40-400x cheaper (w/ output tokens free)”. The homepage examples show a 193.6x speed gain and a 444.6x cost gain. The company says those figures likely sit near the upper end of real-world results, and that the long-term sustainability of the pricing is unproven.

The intended jobs are classification, routing, scoring, extraction, branching, large-dataset processing, real-time applications and guardrails for LLM output. The Robotics Media reports that TypeSafe claims gains of roughly 100x over leading frontier chat models on narrow, high-volume tasks such as classifying support tickets, moderating content or filtering trades. The name comes from Jevons Paradox, the economic observation that greater efficiency can sometimes raise total consumption. TypeSafe wants developers to treat Jev as a software primitive that applications call repeatedly whenever they need a semantic judgment.

TypeSafe’s workflow evaluations run identical compute graphs and compare Jev’s outputs against the average probabilities from GPT-6 Astra and Fable 5.1, with no ground-truth labels. The benchmark multiples rest on only four published workflows, all built by the company’s own model-capabilities team, which leaves room for selection bias. The LLM baseline runs through TypeSafe’s System One LLM wrapper, which constrains conventional models to compatible structured decisions and may affect their speed and cost. Most tests ran from laptops on the US West Coast, where the service is based.

The demos include Doom, which Jev plays from structured game state plus text, with no images, and Wikiracing. TypeSafe notes that its Wikiracing comparisons mostly ran rival LLMs in non-reasoning modes, which makes them look weaker than they might with reasoning enabled.