TypeSafe AI Unveils Jev: A System 1 Model Delivering Typed, Calibrated Decisions Rather Than Plain Text

TypeSafe AI has released Jev, a transformer-based model that breaks from the ChatGPT playbook: instead of generating human-facing text, it answers machine-facing questions. Developers send a state along with typed queries, and Jev responds with typed decisions and probability distributions that code can branch on directly. The model is available now as a hosted API in early access behind a waitlist; TypeSafe has not published weights, a parameter count, or a self-hosting option.

What Is a "System One" Model?

The name borrows from psychologist Daniel Kahneman's distinction between fast intuition and slow, deliberate reasoning. The TypeSafe team argues that RLHF tuned existing models to satisfy human preferences, which produced chat interfaces but also overconfidence and mode dropping—flaws that keep a human in the loop.

Jev runs on an entirely new stack: an undisclosed architecture, a parallel sampler, and a training method called Reinforcement Learning for Calibrated Decisions (RLCD).

How the API Works

A single endpoint—POST https://api.typesafe.ai/v1/systemone—handles everything. The request body carries a state, a model, and a map of questions in three types:

| Type | Purpose | Returns |
|------|---------|---------|
| Choice | Pick one option from a list | choice, probabilities, confidence |
| Score | Rate against ordered levels | score, probabilities, confidence |
| Noul | Judge if a statement is true | noul, a probability from 0 to 1 |

Questions run in parallel and in isolation against the same state, and TypeSafe says adding questions barely affects response time. A Choice supports up to 255 options. The Python SDK installs via `pip install typesafe-sdk` (Python 3.10+), with a JavaScript package at `@typesafe-ai/sdk`, plus cURL support and a Claude Code agent skill in the quickstart.

Confidence Is the Product

Every Choice and Score answer includes a confidence value from 0 to 1, derived from the shape of the probability distribution. In the docs example, "billing" wins at 0.84, but confidence is only 0.596 because "technical" still holds 0.159. The documentation outlines three paths: act automatically on high confidence, review borderline cases, and escalate low-confidence answers to humans—with thresholds scaling to the cost of a wrong action.

Pricing, Speed, and Benchmark Caveats

Jev costs $42 per billion input tokens, against LLMs TypeSafe quotes at $0.20 to $10 per 1M input tokens. In a recorded demo, Jev finished in 0.114s for $0.000081, while GPT-5.6 Terra took 8.566s for $0.013880—figures TypeSafe summarizes as 193.6x faster and 444.6x cheaper.

Those numbers warrant scrutiny, though. The reference answer is the average of GPT-6 Astra and Fable 5.1, TypeSafe's own capabilities team wrote the workflows, the company expects real-world gains at the lower end of that range, and it admits it cannot prove the price is unsubsidized. The "zero hallucinations" claim only means schema matching is guaranteed—not that answers are always correct.

What Developers Are Building

Community projects appeared within days of launch:

• Command safety: Vercel CEO Guillermo Rauch reported Jev up to 18x faster at p95 than GPT Luna with better accuracy, though his fx reviewer still runs on Luna.
• Email triage: Bryo AI CTO Nikhil Mudholkar found Gemini slightly more accurate but 10 to 20 times more expensive.
• Browser agents: Browser Use's jev-ultrafast ran a Zürich-to-London Google Flights search in 7.1 seconds.
• Phone agents: Droidrun's mobile-jev drove Uber on a real Android phone—9 actions in about 21 seconds, with no booking completed.
• Video scoring: jevmeter scores every sentence of a debate for about $0.05.
• Live typing: Steve Krouse's Typewriter updates 16 judgments as you type.
• Games: Jev completed StarCraft's first combat mission and runs the guards in heist-one.
• Agent guardrails: jev-guard rates each tool call as deny, ask, or allow.
• Data and homes: pg-jev adds natural-language filters to Postgres; HA-Jev turns answers into Home Assistant entities.

Key Takeaways

• Jev outputs typed decisions with probabilities, not strings.
• Three primitives (Choice, Score, Noul) can share a single request.
• Input costs $0.042 per 1M tokens; output tokens are free.
• TypeSafe reports 70ms to 500ms end-to-end response times.
• The headline benchmarks are vendor-run, so results should be verified on your own data.

Whether a machine-native decision model can carve out space alongside text-first LLMs will depend on independent benchmarking and how quickly the waitlist opens up.

Meta description: TypeSafe AI's Jev returns typed, calibrated decisions instead of text. Here's how the System One model works, its pricing, and benchmark caveats. (152 characters)

Tags: TypeSafe AI, Jev, System One Model, AI API, machine decision-making

Featured image: Abstract visualization of branching probability nodes and decision pathways in cool blue tones, no people or logos.