Decision Model Cost Lab
If you label, route or score a huge stream of items, a decision (System One) model may cost a fraction of asking a chat model. Set your volume and state size and see the bill for 22 decision models against small chat models. Cost only: we have not measured accuracy.
Lab
The ticket, record or tool call you send for each decision
What a chat model writes back for one label. Decision models write none
10M decisions a month: $50.00 on Perplexity Decider Flash V1.2, against $600 asking OpenAI: GPT-6 Luna for a label.
12x cheaper on price, before you count wrong answers.
Decision models, cheapest first
Cost = state tokens × input price. Output is free on every model listed.
| Model | Input per M | Per 1M decisions | A month | Window |
|---|---|---|---|---|
| Perplexity Decider Flash V1.2 Perplexity | $0.01 | $5.00 | $50.00 | 262K |
| Respan Span-02 Respan | $0.01 | $5.00 | $50.00 | 131K |
| Baseten Joey 1 Baseten | $0.01 | $5.00 | $50.00 | 16K |
| Perplexity Decider V1.2 Perplexity | $0.02 | $10.00 | $100 | 262K |
| Inception Mercury Decide Inception | $0.02 | $10.00 | $100 | 66K |
| Liquid d1 / d1.1 Liquid AI | $0.02 | $10.00 | $100 | 66K |
| Baseten Elephant 1 Baseten | $0.035 | $17.50 | $175 | 16K |
| Cloudflare Clef Flash Cloudflare | $0.038 | $19.00 | $190 | 25K |
| ScaleDown Classify-1 ScaleDown | $0.04 | $20.00 | $200 | 1M |
| Nace.AI Drex v1.5 Nace.AI | $0.04 | $20.00 | $200 | 131K |
| TypeSafe Jev 1.14 / 1.13 TypeSafe | $0.042 | $21.00 | $210 | 262K |
| Liquid d1-ultra Liquid AI | $0.042 | $21.00 | $210 | 33K |
| Microsoft Decision-1 Microsoft | $0.042 | $21.00 | $210 | 33K |
| Together Tev1 4B Experimental Together AI | $0.042 | $21.00 | $210 | 33K |
| Jared Palmer Kev 4B Jared Palmer | $0.042 | $21.00 | $210 | 8K |
| Upstage Solar Decide / Flash Upstage | $0.05 | $25.00 | $250 | 524K |
| Nace.AI Drex Max Nace.AI | $0.06 | $30.00 | $300 | 262K |
| Nace.AI Drex v2.0 Nace.AI | $0.06 | $30.00 | $300 | 131K |
| OpenAI GPT-6 Luna Decisions OpenAI | $0.1 | $50.00 | $500 | 1.05M |
| Baseten Mammoth 1 Baseten | $0.1 | $50.00 | $500 | 16K |
| Cloudflare Clef Omni Cloudflare | $0.15 | $75.00 | $750 | 66K |
| Cloudflare Clef Cloudflare | $0.192 | $96.00 | $960 | 66K |
Prices: OpenRouter live models API, decision models (read 3 November 2026), as published in OverpayingForAI's System One guide. Checked 2026-11-03. Full System One guide.
Small chat models asked for a label
Cost = state tokens × input price + label tokens × output price, at catalogue prices last verified 2026-11-05.
- OpenAI: GPT-6 Luna$60.00 per 1M items$600 a month
- OpenAI: GPT-5.4 Nano$125 per 1M items$1,250 a month
- Google: Gemini 3.5 Flash Lite$200 per 1M items$2,000 a month
Cascade: decision model first, chat model for the unsure ones
Share the decision model is unsure about
- Decision model only
- $210
- Cascade
- $300
- OpenAI: GPT-6 Luna only
- $600
Self-hosting: Open-weight model run on your own GPU (Laya on one always-on T4) costs about $252 a month (OverpayingForAI September 2026 estimate, see /pricing/laya-pricing). On TypeSafe Jev 1.14 / 1.13 that beats paying by the token above roughly 12M decisions a month at this state size. You do 10M.
Cost per correct decision
The cheapest model is not the cheapest answer if it is wrong more often. Test both on a few hundred of your own labelled items, then enter the accuracy you measured. We do not fill these in for you because we have no figure to give.
Optional. Support time, refunds, rework
Enter both accuracies to see the comparison.
What this does and does not tell you
- It compares price only. Output is free on every model listed, so a decision costs its state tokens times the input rate. Accuracy is not in this file: most decision models have no independent test, and we ran none of our own. Models that share a price and window are grouped on one row. Span-01 and Span-01 Lite score agent behaviours, not general states, and are marked as such.
- Chat models can explain themselves and return several fields at once. A decision model returns typed choices with probabilities. If you need free text or reasoning, you need a chat model.
- Some models have a small window. A state longer than the window is greyed out in the table.
Questions people ask
What is a decision model?
A decision (System One) model reads a state, such as a ticket or a JSON record, and returns typed answers with a probability for each option. It does not write text. On OpenRouter its output is free, so a decision costs only the state's input tokens.
Is a decision model cheaper than a small chat model?
On the default settings here, 10M decisions a month at 500 tokens each cost $50.00 on Perplexity Decider Flash V1.2, the cheapest decision model that fits, against $600 asking OpenAI: GPT-6 Luna for a 20-token label. The gap shrinks as the label gets shorter or the state gets longer, and it says nothing about accuracy.
Does this page say which model is most accurate?
No. We ran no accuracy test of our own and most decision models have no independent results. This page compares cost only. Use the cost-per-correct-decision box with accuracy figures from your own labelled sample.
What is a cascade?
A decision model labels every item. Items it is unsure about, below a probability cutoff you set, are sent to a chat model for a second opinion. You pay the decision model on everything and the chat model only on the share you escalate.