Architecture cost review
System One Models: What Decision Models Are and What They Cost
In seven weeks the category went from one model to 33 OpenRouter listings. The prices sit within a few cents of each other per million tokens, so the window size and the evidence behind each model decide more than the rate card.
Direct answer
A System One model, or decision model, reads a state and returns typed answers with probabilities. It writes no text, and on OpenRouter its output is free. At a 500-token state, a million decisions cost $5 on Perplexity's Decider Flash, $21 on Jev and $50 on GPT-6 Luna Decisions. A small LLM writing a 20-token label costs $57 to $125. Use one for routing, gating and triage. Keep an LLM wherever a person reads the answer.
Decision summary
| Decision area | What matters |
|---|---|
| Perplexity Decider Flash V1.2 (+ Flash Latest alias) | $0.01/M input · $5 per 1M decisions · 262,144 tokens |
| Respan Span-02 | $0.01/M · $5 · 131,072 |
| Baseten Joey 1 | $0.01/M · $5 · 16,384 |
| Liquid d1 and d1.1 (+ d1 Latest alias) | $0.02/M · $10 · 65,536 |
| Perplexity Decider V1.2 (+ Decider Latest alias) | $0.02/M · $10 · 262,144 |
| Inception Mercury Decide | $0.02/M · $10 · 65,536 |
| Respan Span-01 / Span-01 Lite (behaviour scorers) | $0.02/M · $10 / free · window not listed |
| Baseten Elephant 1 | $0.035/M · $17.50 · 16,384 |
| Cloudflare Clef Flash | $0.038/M · $19 · 24,576 |
| ScaleDown Classify-1 | $0.04/M · $20 · 1,000,000 |
| Nace.AI Drex v1.5 | $0.04/M · $20 · 131,072 |
| TypeSafe Jev 1.14 and 1.13 (+ Jev Latest alias) | $0.042/M · $21 · 262,144 |
| Liquid d1-ultra (+ Latest alias) | $0.042/M · $21 · 32,768 |
| Microsoft-Decision-1 | $0.042/M · $21 · 32,768 |
| Together Tev1 4B Experimental | $0.042/M · $21 · 32,768 |
| Jared Palmer Kev 4B | $0.042/M · $21 · 8,192 |
| Upstage Solar Decide and Solar Decide Flash | $0.05/M · $25 · 524,288 |
| Nace.AI Drex v2.0 | $0.06/M · $30 · 131,072 |
| Nace.AI Drex Max | $0.06/M · $30 · 262,144 |
| OpenAI GPT-6 Luna Decisions | $0.10/M · $50 · 1,050,000 |
| Baseten Mammoth 1 | $0.10/M · $50 · 16,384 |
| Cloudflare Clef Omni | $0.15/M · $75 · 65,536 |
| Cloudflare Clef | $0.192/M · $96 · 65,537 |
What a System One model returns
You send a state, such as a ticket, a JSON record or a pending tool call, plus typed questions. Each question is yes/no, pick one of several options, or a score on a scale. The model answers in one forward pass and returns a choice with a probability for each option. It never writes a sentence.
TypeSafe coined the term when it launched Jev on 15 September 2026. The name borrows Daniel Kahneman's System 1, the fast, automatic mode of thinking, as opposed to slow, deliberate System 2. A chat model is the System 2 here: it can reason and explain, at a higher price per call.
OpenRouter serves these models through its Decisions API and the System One endpoint, /v1/systemone. OpenAI's own version, GPT-6 Luna Decisions, sits on /v1/decisions. The request shapes differ a little, but the questions you ask carry over.
Reading the price table
OpenRouter lists 33 decision-model entries. Five are ~latest aliases that redirect to a named model, and one is a duplicate free route for Span-01 Lite, which leaves 27 distinct models. The table groups each alias with the model it points to.
Each row shows the input price per million tokens, the cost of a million decisions at a 500-token state, and the context window. The middle number is the input price × 500, because 1 million decisions × 500 tokens is 500 million tokens. For Jev, that is 500 × $0.042 = $21. Output costs nothing on every row.
- Cheapest general models: Decider Flash V1.2 and Span-02 at $5 per million decisions
- Largest windows: GPT-6 Luna Decisions (1,050,000) and Classify-1 (1,000,000)
- Dearest: Cloudflare Clef at $96, more than 19 times Decider Flash
Decision model or small LLM: the same 500-token item
Take one 500-token item and ask a small LLM for a 20-token label. Per million items, the LLM pays for 500 million input tokens and 20 million output tokens. The decision model pays for the 500 million input tokens only. The 500 tokens include the question text on both sides.
GPT-5.4 nano costs $0.20 in and $1.25 out per million tokens: $100 + $25 = $125. Claude Haiku 5.5 and GPT-6 Luna cost $0.10 and $0.50: $50 + $10 = $60 each. Against that, Jev costs $21, Decider Flash $5, Clef Flash $19 and GPT-6 Luna Decisions $50.
So the decision model wins on tokens even against batch pricing. Batch Luna or Haiku 5.5 is $25 + $5 = $30, still above Jev. The decision model also removes the parsing code and returns a probability you can set a threshold on.
- Decider Flash V1.2: $5
- Clef Flash: $19
- Jev 1.14: $21
- GPT-6 Luna Decisions: $50
- GPT-6 Luna or Claude Haiku 5.5: $60 ($30 batch)
- GPT-5.4 nano: $125 ($62.50 batch)
When a decision model is the wrong buy
If a person reads the output, you need text, and these models do not write. The same goes for a reason attached to the decision, a summary, or several free-text fields pulled from a document. An LLM does those in one call. A decision model plus an LLM is two calls.
Long reasoning is the second gap. A decision model answers in one pass. A question that needs several steps of working, such as reconciling two contracts, belongs with a reasoning model. Keep arithmetic and date comparisons in your own code either way.
The third gap is accuracy. On a 77-category intent test cited by Kai Waehner, Jev scored 79.0% against 83.9% and 86.2% for two unnamed OpenAI models, and the article calls the gap significant. Many intents and fine distinctions can still favour a strong LLM. Measure on your own labels before you switch.
Window size changes the ranking
Price per token only matters if your state fits. Kev 4B holds 8,192 tokens, Joey 1 16,384 and Clef Flash 24,576. A 20,000-token email thread fits Clef Flash but not Kev or Joey. A 100,000-token contract fits Jev, both Deciders, all three Drex models, Span-02, Solar Decide, Classify-1 and GPT-6 Luna Decisions.
Long states also make the price gaps real money. A 100,000-token state costs $0.001 per decision on Decider Flash, $0.0042 on Jev and $0.01 on GPT-6 Luna Decisions. At 10,000 such decisions a day, that is $10, $42 or $100 a day.
Where the rest of this cluster goes
For a ranked list of who to buy, with each model's catch, read the best decision models page. For the head-to-heads, read Jev vs Clef (images, latency claims and two independent tests) and Jev vs GPT-5.4 nano (decision model against a small LLM at three item sizes).
The pricing pages cover Jev, Cloudflare Clef, OpenAI's Decisions API and self-hosted Laya one by one. The September pages on Jev vs Laya, Jev vs Kev vs Solar Decide, and whether Jev is worth it still hold for their questions. Their prices and Jev's context window are from 1 October.
Key takeaways
- →A System One model, or decision model, reads a state and returns typed answers with probabilities. It writes no text, and on OpenRouter its output is free. At a 500-token state, a million decisions cost $5 on Perplexity's Decider Flash, $21 on Jev and $50 on GPT-6 Luna Decisions. A small LLM writing a 20-token label costs $57 to $125. Use one for routing, gating and triage. Keep an LLM wherever a person reads the answer.
- →Find the two decision points in your app that run most often, price them with the table below at your real state size, and test two decision models against your current LLM call on a few hundred labelled cases.
- →Most accuracy claims in this field come from the vendors; the few independent tests we cite use small samples and older price points.
What users report
Users who moved high-volume classification calls off LLMs report big cost and speed gains; skeptics say plain LLMs or trained classifiers already did this job.
Paraphrased from public posts, with the date and a link to each original. These are other people's accounts, not our tests, and we have not verified each one.
One user reported that Jev made several workloads previously run on Luna or Gemini "10x cheaper and 2x faster", by their own measurement.
Value · Jev · Hacker News, Sep 2026
A developer reported moving about 20 unreliable Luna judgement calls (tool choice, topic routing) to Jev in a couple of hours; their own numbers went from roughly 70% correct to near 100%, with 800ms median latency.
Switching · Jev · Hacker News, Oct 2026
One user said they use Jev-style models only for straightforward decisions, in 300 ms or less at $0.042 per million input tokens with output unbilled, and not for high-stakes calls.
Value · Jev · Hacker News, Oct 2026
A machine learning engineer argued the hype was overblown, saying plain LLMs had already been "perfectly cheap and serviceable" zero-shot classifiers for some time.
Value · Decision models · Hacker News, Oct 2026
One user tested Jev on OpenRouter and found it only "okay", reporting that a task-specific classifier was significantly better, free to run, and faster.
Quality · Jev · Hacker News, Sep 2026
How this page was prepared
Prices on this page come from OpenRouter's live models API, read on 3 November 2026. Cost per decision is state tokens multiplied by the input rate, because output is free on every decision model listed; small-LLM costs add the stated label length at that model's output rate. OpenRouter's credit fee is shown separately where it matters. We ran no private benchmark for these pages: accuracy and latency figures are credited to the vendor or third party that measured them, and vendor figures are labelled as vendor claims.
- OpenRouter live models API (checked 3 Nov 2026) ↗
- OpenRouter live models API, all models (checked 3 Nov 2026) ↗
- OpenRouter — Jev (TypeSafe) documentation hub ↗
- Simon Willison — Jev and System One decision models ↗
- Kai Waehner — How System One models like Jev change enterprise AI architecture (28 Sep 2026) ↗
- DEV Community (ankushkun) — Clef beat Jev on Jev's home turf ↗
Frequently asked questions
What is a System One model?
A model that reads a state and answers typed questions (yes/no, pick one, or a score) with probabilities, in one forward pass, without writing text. TypeSafe coined the name with Jev on 15 September 2026, after Kahneman's fast System 1 thinking.
Is a decision model the same as a classifier?
Close. A classic classifier is trained for one fixed label set. A decision model takes the questions and options in the request, so one model can route tickets, gate tool calls and score urgency without retraining.
Why is the output free?
On OpenRouter every decision-model listing shows $0 for output on 3 November 2026. The answer is a choice and a few probabilities, so the bill is set by the state you send.
Which decision model is cheapest?
Among general models, Perplexity's Decider Flash V1.2 and Respan's Span-02 at $0.01 per million input tokens, or $5 per million 500-token decisions. Span-01 Lite is free but only scores behaviours in a conversation.
Can I send images to a decision model?
Some take them. OpenRouter's listings say Span-02, d1-ultra, Drex Max and GPT-6 Luna Decisions accept images, and Clef takes up to four per request. Jev is text and JSON only.
How do I know a decision model is accurate enough?
Export a few hundred past cases with known answers, run them through the model in shadow mode, and pick a probability threshold from the results. Few public tests exist, and most figures are the vendor's own.