Cua Releases CUA-S1 Form-Decision Model as TypeSafe AI Launches Jev System One API

In the Hacker News post, Cua founders Dillon and Francesco said they wanted to test how many computer-use tasks need a full general-purpose large language model to plan, explore paths, and recover from failure, and how many are local decisions such as placing a value in a box, checking a box, or ignoring an element. They cited TypeSafe's Jev and its System One Model framing, which borrows Daniel Kahneman's distinction between fast, intuitive thinking and slow, analytical reasoning. CUA-S1-FORMS was built from ideas and code in jevlike, and then a second model was trained for form interactions. It has 706,000 parameters and the original checkpoint is 2.8 MB. The first training iteration took less than 30 minutes on synthetic data.

Given structured elements and values extracted from a document, CUA-S1-FORMS predicts whether to use the given value, CHECK, CLICK, or SKIP for each element. It does not predict new values for text fields and does not consider screenshots. Element decisions are scored together, and Cua said a developer's code can order the actions while Cua Driver executes them one at a time.

Cua reported a first evaluation of this specialist against hosted Jev on its form task. For the whole decision set, CUA-S1-FORMS was 99.7 percent correct versus 83.6 percent for hosted Jev. For steps requiring an action, it was 100 percent correct versus 96 percent. For steps that leave already-filled fields alone, it was 100 percent correct versus 74 percent. Cua noted that the specialist was trained specifically for this task and convention, while hosted Jev has not been fine-tuned for it, making the comparison an experiment in scoped specialization. Cua measured 7 to 9 milliseconds to score a form locally versus 260 to 280 milliseconds per call to hosted Jev including network latency, but said the samples measure different things and are not end-to-end form completion times. Cua open sourced synthetic data generation, training, evaluation, and Driver integration under libs/cua-s1 with an MIT license.

MarkTechPost reported that TypeSafe AI's Jev is transformer-based but is not a large language model. It does not generate text; instead, users send a state and typed questions, and it returns typed decisions with probabilities that code can branch on. Jev is available as a hosted API in early access behind a waitlist. TypeSafe has not published weights, a parameter count, or a self-hosting option. TypeSafe argues that RLHF-tuned models optimized for human preference produced chat, overconfidence, and mode dropping, flaws that keep a human in the loop. Jev uses a new stack that TypeSafe describes as a new architecture, a parallel sampler, and Reinforcement Learning for Calibrated Decisions, or RLCD. The architecture has not been disclosed.

The Jev API uses one endpoint, POST https://api.typesafe.ai/v1/systemone. The body carries state, model, and a map of questions. The documentation defines three question types: Choice, which picks one option from a list and returns choice, probabilities, and confidenceScore; Score, which rates against ordered levels and returns score, probabilities, and confidence; and Noul, which asks whether a statement is true and returns a probability from 0 to 1. Questions run in parallel and in isolation against the same state, and TypeSafe says adding questions barely changes response time. A Choice supports up to 255 options. Python and JavaScript SDKs are available.

Confidence is central to Jev's output. Every Choice and Score answer carries a confidence value from 0 to 1, derived from the shape of the probability distribution. In a TypeSafe documentation example, billing wins at 0.84 but confidence is only 0.596 because technical still holds 0.159. The docs suggest acting on high confidence, reviewing the middle, and sending low confidence to a human, with thresholds scaled to the cost of a wrong action.

TypeSafe prices Jev at $42 per billion input tokens, with output tokens free, equivalent to $0.042 per 1 million input tokens. MarkTechPost reported that TypeSafe quotes existing LLMs at $0.20 to $10 per 1 million input tokens. In TypeSafe's recorded demo, Jev finished in 0.114 seconds for $0.000081, while GPT-5.6 Terra took 8.566 seconds for $0.013880. TypeSafe claims Jev is 193.6 times faster and 444.6 times cheaper. MarkTechPost noted caveats: the figures come from TypeSafe's own workflow evaluations, the reference answer is the average of GPT-6 Astra and Fable 5.1, TypeSafe's capabilities team wrote the workflows, TypeSafe expects the gains to sit at the high end of real use, and TypeSafe says it cannot prove the price is unsubsidized. TypeSafe's 'zero hallucinations' claim means schema matching is guaranteed; the 0 percent figure is not empirical, and answers can still be wrong.

Community projects appeared within days of Jev's launch, according to MarkTechPost. Vercel CEO Guillermo Rauch reported that Jev was up to 18 times faster at p95 than GPT Luna and more accurate, while an fx reviewer still ran on Luna, with engineer Pranit Sharma sharing the benchmark. Bryo AI CTO Nikhil Mudholkar found Gemini slightly more accurate for email triage but 10 to 20 times more expensive. Browser Use's jev-ultrafast ran a Zürich to London Google Flights search in 7.1 seconds. Droidrun's mobile-jev drove Uber on a real Android phone with 9 actions in about 21 seconds, though no booking was completed. Other projects include jevmeter, which scores every sentence of a debate for about $0.05; Steve Krouse's Typewriter, which updates 16 judgments as you type; Jev completing StarCraft's first combat mission; jev-guard, which rates tool calls as deny, ask, or allow; pg-jev for Postgres filters; and HA-Jev for Home Assistant entities.

Cua said its interest is the space between a brittle script and a general agent loop: form fields vary enough that scripts become unwieldy, but the available decisions can remain narrow and well scoped. The company wants to explore a general agent encountering something novel and passing well-understood decisions to specialists. The current CUA-S1 release is for forms only.