TypeSafe opens Jev early access for fast, typed AI decisions

Why it matters

Jev tests whether AI automation needs narrower, cheaper decision models rather than ever-larger chat systems. TypeSafe's case is credible enough to test, but its benchmark claims still come from TypeSafe.

Diogo Almeida (@CompleteSkeptic) opened early access to TypeSafe AI's Jev model, betting that software automation needs fast, constrained decisions rather than another system trained to write fluent answers.

https://x.com/CompleteSkeptic/status/2099925682726002904

Almeida, TypeSafe's co-founder and CEO, presented Jev in a thread on X on September 15th after TypeSafe published its launch materials on September 14th. He spent the previous two years building TypeSafe with co-founders Sasha Sheng, a former Meta and FAIR research engineer, and Erik Gafni, who previously built genomics AI startup Ravel and worked at Invitae and Freenome.

Almeida's break with chat models carries some history. He co-authored OpenAI's 2022 InstructGPT paper, which documented the human-feedback training process that helped turn GPT-3 into a more useful instruction-following system. His name also appears in OpenAI's credits for the original ChatGPT release, and he previously worked at Google Brain. TypeSafe's thesis is a critique of the path that work helped establish: models optimized to produce responses people prefer still require supervision when their output controls software.

Jev takes a narrower assignment. Developers send the model a block of state and a set of typed questions. Jev returns choices, scores or probabilities that application code can use to classify, route, approve or escalate work. It does not generate prose, code or open-ended strings.

TypeSafe calls Jev the first "System One" model, borrowing the name from the fast, intuitive mode of thought described by Daniel Kahneman. TypeSafe says Jev uses a new architecture, a parallel sampler and a training method called Reinforcement Learning for Calibrated Decisions, or RLCD. The release describes the approach at a high level rather than publishing model weights or a research paper with enough detail for outside researchers to reproduce the results.

The speed comes from giving up strings

Traditional large language models generate tokens sequentially, with each token conditioned on those that came before it. TypeSafe says Jev evaluates multiple typed questions in parallel and returns their probability distributions in one response. According to TypeSafe's developer documentation, adding questions has little effect on response time because each judgment is evaluated independently against the same input state.

That design sets clear boundaries around what Jev can do. TypeSafe recommends decomposing a broad task, such as assessing a startup pitch, into separate questions about market size, feasibility and differentiation. Developers then encode the weighting and final action in ordinary software. Jev supplies the fuzzy judgments; code keeps control of the workflow.

TypeSafe prices Jev at $42 per billion input tokens, or $0.042 per million, and says output is free because the architecture makes it too inexpensive to meter. TypeSafe reports end-to-end response times between 70 milliseconds and 500 milliseconds, with gains of 40x to 200x against conventional models on tasks shaped for Jev.

The largest figures require more caution. TypeSafe's homepage advertises one workflow result as 193.6x faster and 444.6x cheaper. TypeSafe says those numbers sit near the high end of the gains customers should expect. Its published workflow evaluations cover four examples: security incidents, agent-trace monitoring, invoice processing and customer service.

Those evaluations are internal. TypeSafe's capabilities staff created the workflows, and reference answers come from the average output of GPT-6 Astra and Claude Fable 5.1 rather than an independently established ground truth. TypeSafe acknowledges that its staff's role may introduce bias. No independent evaluation accompanied the launch.

The "zero hallucinations" claim also uses a narrower definition than developers usually attach to the term. TypeSafe guarantees that Jev's response will match the requested schema, preventing invented fields, malformed tool calls and invalid output types. Jev can still choose the wrong answer. The probability and confidence fields are intended to let software set thresholds for autonomous action and route uncertain cases to review.

A model built to disappear inside software

TypeSafe demonstrated Jev making roughly 10 calls a second to control a Doom bot for about $7 an hour. The bot receives a structured, text-based representation of the game state rather than images. Another demo asks Jev to navigate between Wikipedia pages by selecting among large sets of links. Jev currently supports as many as 255 choices, TypeSafe says.

The demonstrations show the product TypeSafe is trying to sell: machine intelligence that functions like an inexpensive software primitive instead of a visible assistant. That puts Jev closer to a probabilistic decision engine than a general chatbot, with usefulness tied to how well developers break real business processes into narrow judgments and deterministic rules.

Early access also limits how broadly TypeSafe's performance claims can be tested. TypeSafe is bringing developers off a waitlist rather than offering unrestricted access. The architecture's commercial case will rest on production workloads where latency and repeated model calls matter enough to justify rebuilding workflows around Jev's constrained interface.

Almeida named the model after Jevons paradox, the economic observation that efficiency gains can increase total consumption. His wager is that a large reduction in inference cost will make developers put model calls into places where today's LLM economics make little sense. Jev gives up the feature that made generative AI famous to chase that volume: it cannot write the answer, only decide what software should do next.