AI ExplainersNov 9, 20269 min read

What Is a System One Model? An AI That Picks, Never Writes

A System One model picks one option from your list and writes nothing. On DecideBench v2 the best got 94.4% right at about 1/290 of a frontier LLM's cost.

On this page
  1. System One model meaning: where does the name come from?
  2. System one vs system two models: what is the difference?
  3. System one model vs LLM
  4. System One model architecture: how does one work?
  5. What System One models are good at, and where they fail
  6. Examples: a list of System One models
  7. When should you use a System One model?
  8. FAQ
  9. Sources

A System One model is an AI model that picks instead of writing. You give it an input, a question and a short list of options, and it returns one option key, usually with a probability for each. On 9 November 2026, DecideBench v2 measured 26 of them on 450 decisions. The best, Nace.AI's Drex Max, got 94.4% right for about 1/290 of a frontier LLM's cost.

Key facts

  • What it returns: one of the options you supplied, such as refund, escalate or rollback, with probabilities, never free text.
  • Who named it: TypeSafe, when it launched Jev on 15 September 2026.
  • How good the best ones are: Drex Max got 94 of every 100 decisions right on DecideBench v2. Claude Opus 4.8, the frontier LLM the test was built against, got 99.3%.
  • What it costs: Drex Max came to $91.00 per million decisions. Opus 4.8 came to $26,542.70 per million.
  • Where they slip: decisions that need arithmetic checked against an ordered list of rules. On staged software releases, no System One model scored above 74%.

DecideBench is our benchmark. The data, code and every model's price source are public on GitHub.

System One model meaning: where does the name come from?

The name comes from Daniel Kahneman's split between fast, intuitive thinking (System 1) and slow, deliberate thinking (System 2). TypeSafe coined "System One models" in its Jev launch post on 15 September 2026. Its definition: "a new class of frontier models built to make fast, structured decisions that software can use directly." It says the name "draws on the distinction between fast, intuitive System 1 thinking" and "slow, deliberate System 2 reasoning", citing *Thinking, Fast and Slow*.

Six days later Simon Willison wrote it up as "a new shape of LLM": "unstructured state in, typed probabilistic decisions out." He also used the other name the category goes by, decision models. The two terms mean the same thing, and this post uses both.

If you searched for a System One model paper: TypeSafe's launch post is where the term starts, and it cites Kahneman's book rather than a research paper of its own. The page gives no parameter count and describes only "a new model architecture" and a "parallel sampler".

System one vs system two models: what is the difference?

A System One model answers directly. A System Two model reasons step by step before it answers. In practice, System Two means a general LLM with thinking turned on, and DecideBench uses Claude Opus 4.8 with adaptive thinking as its System Two reference.

The same 450 decisions, both ways:

Claude Opus 4.8 (System Two reference)Drex Max (best System One model)
Right answers99.3% (447 of 450)94.4% (425 of 450)
Both halves of a pair right98.7%89.3%
Cost per million decisions$26,542.70$91.00
Median time per decision1,336 ms258 ms

"Both halves of a pair right" needs one line of explanation. DecideBench writes its items in pairs that differ by one small, realistic edit that changes the right answer, such as a sponsorship pitch instead of a bug report from the same customer. A pair counts only if the model gets both halves right, which catches models that answer from the topic instead of the detail.

The trade is plain in the table. Opus 4.8 got 22 more decisions right out of 450. Drex Max was about five times faster and about 290 times cheaper. Whether those 22 decisions are worth the difference depends on what a wrong one costs you.

System one model vs LLM

A System One model is built on the same kind of network as an LLM, but it gives up writing. A general LLM can answer any question in any words, and your software then has to check those words. A System One model can only return an option you defined before the call, so there is nothing to parse and no invented fourth answer.

Returning a fixed option is also why System One models are fast. A chat model writes its reply one token at a time. TypeSafe says Jev "outputs all probabilities in parallel instead of autoregressively generating by token." Pricing follows too: Drex Max lists $0.06 per million input tokens on Nace.AI's page, and output is free because there is no output text to bill.

We compared the two head to head in decision model vs LLM, including where the LLM earned its price.

System One model architecture: how does one work?

Most System One models we tested are small LLMs fine-tuned to score a list of options. You send the input, the question and the options. The model returns a probability for each option, and the highest one is the answer. Their names give their sizes. The smallest have 4 billion parameters, like yev1-4b. The largest have 26 or 27 billion, like REAX 1 and pplx-decider.

DecideBench also tested a second shape: small encoder models that score options without being an LLM at all. They were close to guessing.

Right answers on DecideBench v2's 450 decisions
Right answers on DecideBench v2's 450 decisions
Right answers
Drex Max (hosted)94.4%
pplx-decider v1.2 27B (open)92.2%
Jev (hosted)86.2%
yev1-4b (open)79.3%
GLiNER2.5-Decide 340M (encoder)40.7%
Laya typed-decisions 421M (encoder)39.1%
Julia-1 144M (encoder)26.9%
Random guessing25.1%

Source: DecideBench v2, measured 9 November 2026

Random guessing scores 25.1% because most items have three to six options. The encoders landed at 27–41%. A fourth near-chance entry, CLM-v0.1-8B, scores options through an 8-billion-parameter Qwen model and got 26.4%. Laya, Julia-1 and CLM ran zero-shot, without the one solved example per option that every other model saw. Laya, Julia-1 and CLM were also the cheapest and fastest entries in the test. On long production-style inputs, they did not read the case.

The size of the model mattered less than how it was trained. yev1-4b, at 4 billion parameters, scored 79.3%. Kev-9B, more than twice its size, scored 35.8%.

What System One models are good at, and where they fail

System One models are good at reading intent and routing work, and weakest when the answer depends on arithmetic checked against a list of rules. Here is Drex Max, the best of them, family by family, beside Opus 4.8:

The decisionOpus 4.8Drex MaxJev
Approving an AI agent's proposed action100%100%94%
Routing an agent to a tool98%96%94%
Checking a claim against its sources98%98%94%
Moderating a post100%100%96%
Triaging an inbound email98%98%88%
Ruling on a return100%90%76%
Gating a staged software release100%74%60%
Reading a customer's intent100%100%96%
Triaging a bug report100%94%78%

Each family has 50 items, so a single family number moves by 2 points per decision.

The release gate is the clearest gap. Each item gives a release controller's metrics and an ordered list of written rules: compute a crash-free rate or an error ratio, then apply the first rule that fits. Opus 4.8 got all 50 right. No System One model got more than 37.

The per-option breakdown shows why: every model in the top ten got the "pull the update" answer wrong on all three items that needed it, and Drex Max got only three of the six items where rollback was the right call. Ruling on a return, which mixes dates, amounts and a written policy, is the next widest gap.

On reading a customer's intent, moderating a post or approving an agent's action, the best System One models matched the LLM.

Examples: a list of System One models

The table lists the ten most accurate System One models on DecideBench v2. Hosted ones are called through the vendor's API; open ones you run on your own GPU, so the data stays on your hardware.

ModelMakerHow you get itRight answersCost per million
Drex MaxNace.AIHosted API94.4%$91.00
REAX 1 27BDecision ModelsHosted API92.4%$120.29
pplx-decider v1.2 27BPerplexityOpen weights92.2%$386.34
s1-pro 26BDecision ModelsHosted API91.6%$120.93
JevTypeSafeHosted API86.2%$75.04
Clef 27BCloudflareOpen weights85.3%$216.42
GPT-6 LunaOpenAI Decisions APIHosted API84.0%$140.01
s1-fast 26B-A4BDecision ModelsHosted API82.0%$29.82
yev1-4bXY SpaceOpen weights79.3%$71.15
OpenJev Flash 9BOpenJevOpen weights75.3%$41.33

yev1-4b is our own model, from XY Space. Self-hosted cost is GPU time on one NVIDIA RTX PRO 6000 (AWS g7e.2xlarge, $3.36 an hour on demand) at four requests in flight, and self-hosted latency was measured on that machine with no network in it. We called Jev through AI Space, XY Space's platform.

The full 26, ranked by reader need, are in the best System One models of 2026. The ones you can download are in open source System One models, and how the test works is in the System One model benchmark.

Some System One models are not in DecideBench v2: Liquid AI's d1-ultra, Microsoft-Decision-1, Strands Decider 2B, Upstage Solar Decide and meraGPT Decider 1. We have no score for them.

When should you use a System One model?

Use one when the same decision runs thousands of times, the options are fixed, and cost or latency is the limit. Routing support email, tagging inbound tickets and approving routine agent actions all fit. At $91 per million decisions, a million emails a month costs less than one Opus 4.8 run of 3,500 decisions ($26,542.70 per million works out to about $93 for 3,500).

Keep an LLM for the hard cases rather than choosing one or the other. Send every decision to the System One model first, accept the confident answers, and pass the low-probability ones to a reasoning model. That pattern is a cascade, and it keeps most of the cost saving while the LLM handles the release gates and returns that need arithmetic.

Before you trust any model's probabilities on your own data, test it on a few hundred of your own labelled cases. How to test a classifier covers the method.

XY Space builds agentic workflows that use decision models, and AI Space hosts Jev and open decision models on Western infrastructure.

FAQ

Is Jev a System One model?

Yes. Jev is the model TypeSafe launched the term with, on 15 September 2026. It answers yes/no, pick-one and score questions about a piece of text with probabilities, and writes no text. On DecideBench v2, Jev got 86.2% of 450 decisions right at $75.04 per million and 334 ms. Our Jev explainer covers it in full.

Are System One models LLMs?

Most are built from LLMs. The top open ones are language models with 4 to 27 billion parameters, fine-tuned to score a list of options instead of writing text. Simon Willison called Jev "a new shape of LLM". The difference is the output: a probability per option, never a reply. A few System One models are small encoders instead, and those scored close to random on DecideBench v2.

Can a System One model explain its answer?

No. It returns an option and its probabilities, with no reasoning text. The probabilities tell you how confident it is, which is useful for deciding what to check, but not why it chose. If you need a written reason, send the case to an LLM as well, or only the low-confidence cases, as in a cascade.

Is a System One model the same as a classifier?

Close, with one difference. A classic classifier is trained on a fixed set of labels and needs retraining to add one. A System One model takes the question and the options at call time, so the same model can route email in one call and gate a release in the next. You still test it the way you test a classifier: on your own labelled data, per class. See how to test a classifier.

Sources

Written by

Cho Yin Yong

Principal AI Solutions Engineer, XY Space

Principal AI Solutions Engineer at XY Space. University of Toronto lecturer for five years, co-author of two patents, winner of two competitive AI awards, and nine years of regulated engineering leadership.

More from Cho Yin Yong

Share this article

Work with us

We build the systems these posts describe, and we'll tell you in the first call whether yours is worth building.

Start a project
Work with us

Book a call.We'll come back with specifics.

Start with the map of your organization, or with the one job that hurts. Measured in hours and money, and everything we build stays yours.

Loading form…