TypeSafe AI Releases Jev Typed Decision Model

TypeSafe AI Releases Jev, a transformer based model that does not write a single word. Instead of returning text, it returns typed decisions with probabilities attached, so software can branch on the answer without a human reading it first. The company opened early access through a hosted API and a waitlist.

The pitch is narrow on purpose. You send a state and a set of typed questions. You get back typed decisions. No prose. No hedging paragraph. No parsing layer bolted on afterward to figure out what the model meant.

What Jev Actually Returns

One endpoint handles the work: POST https://api.typesafe.ai/v1/systemone. The request body carries a state, a model name, and a map of questions. Those questions run in parallel and in isolation against the same state, which is why TypeSafe says adding questions barely moves the response time.

The docs define three question types. Choice supports up to 255 options. And Noul answers a yes or no style judgment, which is a simple kind of call that doesn't need numbers. Score gives a numeric rating. It's short. A short Python example shows the shape of it:

from typesafe_sdk import Choice, Noul, TypeSafeClient
client = TypeSafeClient()
r = client.system_one(
state=ticket,
questions={
"department": Choice(instructions="Which team should handle this", criteria={"billing": "Payment issues", "technical": "Bugs"}),
"is_urgent": Noul(instructions="The message conveys urgency"),
},
)
print(r.answers["department"].choice, r.answers["is_urgent"].noul)

Installation is pip install typesafe-sdk on Python 3.10 or later. A JavaScript SDK ships as @typesafe-ai/sdk. The quickstart also covers cURL and an agent skill for Claude Code.

Confidence is the product here, not the label. Every Choice and Score answer carries a value between 0 and 1, derived from the shape of the probability distribution. In the docs example, billing wins at 0.84, but confidence sits at only 0.596 because technical still holds 0.159 of the mass. The suggested paths are simple: act on high confidence, review the middle, send low confidence to a human. Thresholds should scale with the cost of a wrong action.

The System One Bet

The name borrows from Daniel Kahneman's split between fast intuition and slow reasoning. The TypeSafe team argues that RLHF tuned models for human preference, which produced chat, overconfidence, and mode dropping. Those flaws are exactly why a human stays in the loop. Jev uses a different stack: a new architecture, a parallel sampler, and Reinforcement Learning for Calibrated Decisions, or RLCD. TypeSafe has not disclosed the architecture. It has also not published weights, a parameter count, or a self-hosting option.

TypeSafe AI released Jev into a market already crowded with inference providers, and that's a fact you can't ignore.

But here's the thing. Jev isn't competing on chat quality. It's competing on something else entirely, on whether code can actually trust the output shape, which is a question that most providers in this crowded space don't even bother to ask because they're too busy chasing chat benchmarks. So that's the difference.

Pricing, Speed, and the Fine Print

Jev costs $42 per billion input tokens. TypeSafe quotes existing LLMs at $0.20 to $10 per 1M input tokens. In a recorded demo, Jev finished in 0.114 seconds for $0.000081. GPT-5.6 Terra took 8.566 seconds for $0.013880. The team claims 193.6x faster and 444.6x cheaper.

Those figures come from TypeSafe's own workflow evals. 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 these gains to sit at the high end of real use, and says it cannot prove the price is unsubsidized.

Then there is the "zero hallucinations" line. It means schema matching is guaranteed. The 0% figure is not empirical. Answers can still be wrong.

TypeSafe reports 70ms to 500ms end-to-end response times. Output tokens are free. The main benchmarks are vendor run, so the practical advice is the obvious one: test on your own data.

What Developers Built in Days

Community projects appeared within days of launch, and the range is wider than a single use case.

• Command safety: Vercel CEO Guillermo Rauch reported Jev up to 18x faster at p95 than GPT Luna, and more accurate. His post noted the fx reviewer still ran on Luna. Engineer Pranit Sharma shared the benchmark.
• 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 with 9 actions in about 21 seconds. No booking was 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.
• Agent guardrails: jev-guard rates each tool call as deny, ask, or allow.
• Data and homes: pg-jev adds plain language filters to Postgres, while HA-Jev turns answers into Home Assistant entities.

Two of those deserve a second look. Mudholkar's email triage result is the most honest datapoint in the set, because it admits Gemini was slightly more accurate. He still went with Jev. The math changed the decision. Droidrun's phone agent is the other one. Nine actions on a real Android device sounds impressive until you notice the sentence that follows: no booking was completed. Speed is not completion.

The Catch Nobody Should Skip

TypeSafe AI releases Jev as an early access product behind a waitlist. No weights. No self hosting. That means every typed decision your application depends on routes through one company's API, and for developers who spent the last two years building fallback chains around model deprecations, that's a familiar tradeoff they've already learned to live with.

The typed output does solve a real problem. Parsing free text into structured decisions is where a lot of agent reliability goes to die. When the model returns a Choice with a confidence score, the branch logic writes itself. Whether that is worth $42 per billion input tokens depends entirely on what a wrong decision costs you.

Three takeaways matter. Jev outputs typed decisions with probabilities, not strings. Three primitives, Choice, Score, and Noul, can share one request. And the headline benchmarks are vendor run, so the only benchmark that counts is the one you run yourself.

Frequently Asked Questions

What does Jev return instead of text, and how does that change how software uses it?

Instead of returning text, Jev returns typed decisions with probabilities attached. This lets software branch on the answer without a human reading it first. The company opened early access through a hosted API and a waitlist.

How does the single POST endpoint handle multiple typed questions against the same state?

The request body carries a state, a model name, and a map of questions, and those questions run in parallel and in isolation against the same state. TypeSafe says adding questions barely moves the response time because of this parallel execution. The docs define three question types: Choice, Score, and Noul.

Why does TypeSafe argue that RLHF-tuned models are a poor fit for code that depends on typed decisions?

The TypeSafe team argues that RLHF tuned models for human preference, which produced chat, overconfidence, and mode dropping, and those flaws are exactly why a human stays in the loop. Jev instead uses a different stack: a new architecture, a parallel sampler, and Reinforcement Learning for Calibrated Decisions, or RLCD. TypeSafe has not disclosed the architecture and has not published weights, a parameter count, or a self-hosting option.

What do the recorded demo figures show about Jev's speed and cost compared with GPT-5.6 Terra?

In a recorded demo, Jev finished in 0.114 seconds for $0.000081, while GPT-5.6 Terra took 8.566 seconds for $0.013880. The team claims 193.6x faster and 444.6x cheaper, though those figures come from TypeSafe's own workflow evals. TypeSafe says it cannot prove the price is unsubsidized and expects the gains to sit at the high end of real use.

What did early community projects reveal about Jev's real-world performance and limitations?

Community projects appeared within days, including email triage where Bryo AI CTO Nikhil Mudholkar found Gemini slightly more accurate but 10 to 20 times more expensive, and still went with Jev because the math changed the decision. Droidrun's mobile-jev drove Uber on a real Android phone with 9 actions in about 21 seconds, but no booking was completed, showing that speed is not completion. The article notes the main benchmarks are vendor run, so the only benchmark that counts is the one you run yourself.