TL;DR (for anyone paying a chat model to answer multiple-choice questions) - think of a System One model as a fast, cheap if-statement that knows how sure it is.
What is it? 💡
Given a state and a typed question, you get back JSON with a probability for each possible answer.
It works like a multiple-choice exam: you write the question and a group of possible answers, and the model picks the most likely one, noting each one's probability of being right. It is built for routing, classification and scoring at scale, where running a full chat model on every item is slow and expensive.
There are three question types: Choice picks an option from a set, Score rates on a scale you define, and Noul returns how likely a statement is to be true.
System One is Kahneman's fast, instinctive thinking, as opposed to slow, deliberate System Two (like Fable 5.1). Jev (the new model everyone is hyped about, created by TypeSafe AI) is a reference to Jevons.
response = client.system_one(
state="Hi, I've been trying to connect my Stripe account for 3 days and "
"the integration keeps failing. I'm losing sales. Please help ASAP.",
questions={
"department": Choice(
instructions="Which team should handle this",
criteria={
"billing": "Payment or subscription issues",
"technical": "Bugs or integration problems",
"sales": "Pricing or account questions",
},
),
"frustration": Score(
instructions="How frustrated the customer appears",
criteria=[
"Calm, just stating facts",
"Frustrated but civil",
"Very angry, strong language",
],
),
"is_urgent": Noul(
instructions="The message conveys urgency or time-sensitivity",
),
},
)
response.answers["department"].choice # "technical" (0.85, billing 0.15)
response.answers["frustration"].score # 1.0, "Frustrated but civil"
response.answers["is_urgent"].noul # 1.0
Example and results from TypeSafe's quickstart.
Q&A 💬
How was it trained? In two rounds, like a chat model. First, a language model learns from a huge amount of text. Then a second round shapes how it behaves. For chat models, that round (RLHF) rewards answers people like. TypeSafe's version, RLCD, rewards honest probabilities: when Jev says 0.8, it should be right about 80% of the time. The training data is entirely synthetic, a bet the founder rates above RLHF.
Isn't this just a zero-shot classifier? That was a common pushback on Hacker News. Classifiers have done this for years with no training, and Outlines can force open models into a schema. Jev claims two things are new: LLM-level accuracy and probabilities you can trust (while being cheap and fast).
Can it hallucinate? It cannot return a value outside your schema, but it can pick the wrong one.
So is the confidence number trustworthy? Usually when it is very high. It is worked out from the answer probabilities: one clear winner means high confidence, a close race means low.
Thoughts 🧠
- ✅ Why use - When your code asks an LLM the same kind of closed question thousands of times: route this ticket, label this email, is this spam. You get a probability with each answer, so you can automate the sure ones and send the rest for review. It answers fast in 70 to 500ms by TypeSafe's numbers, and JevBench, an independent benchmark, puts it 6 to 7x cheaper per decision than GPT-5.6 Luna or Gemini 3.1 Flash-Lite.
- 🙅 Why not - When the answer isn't on a fixed list, or you need text back: that is still a chat model's job. When accuracy matters more than cost: on JevBench, GPT-5.6 Luna is more accurate and better calibrated. When messages can be off-topic: Jev won't say "none of these" unless you offer it. Zero data retention is enterprise-only, while a free open model you can host yourself scores within one point of Jev.
- ⚔️ Players - Jev (hosted, output tokens free), SemIf (MIT, reads logits off a frozen Qwen3.5-4B, second on JevBench), Laya (Apache-2.0, 421M, cheapest per decision on JevBench), Kev (Jared Palmer, Apache-2.0, 0.8B to 9B on Qwen3.5, speaks Jev's API), GLiNER2 (Apache-2.0, runs on CPU, predates the launch, last of 24 on JevBench), and Outlines (Apache-2.0, schema-constrained decoding).
- 🔮 Prediction - It seems like calibration is going to be the hard part of using these models. 🤠 I feel like this is really early, still waiting to see more actual production use cases that make this make sense. I do think if we wait a little the big shops will have their own similar model (looking at OpenAI/Anthropic/Google etc.) I also see harnesses in the coming weeks implementing some of Jev's ideas, for better accuracy and price optimization.
Flat near 1 until the September 15 launch. Source: Google Trends.
Further reading 📚
- Introducing System One Models & Jev - the launch post. The receipts section admits the headline 444.6x figure sits "on the higher end of real world gains".
- TypeSafe's smart home demo - a six-minute Loom, and the best explanation I found of how you would actually use it.
- Jev 1.13 jaggedness - nine named failure modes with worked fixes.
- The launch thread on Hacker News - the founder answering skeptics directly in the replies.
Thanks to @TomGranot, who edits every issue and has never once returned a low-confidence answer. Tom's note: a wise copywriter once told me that writing is a series of hallucinations, and editing is a series of justifications of hallucinations to make them feel real. I feel like that saying is apt here.