AI ComparisonsNov 9, 20269 min read

System One Model Pricing: What a Million Decisions Costs

System One model pricing is input only: a million decisions cost $29.82 to $140.01 on hosted APIs, against $26,542.70 for Claude Opus 4.8.

On this page
  1. What do System One models charge per token?
  2. Why is the output free?
  3. What does a million decisions cost?
  4. Which model is cheapest at the accuracy you need?
  5. How much of the bill is worked examples?
  6. How is self-hosted pricing worked out?
  7. What does a realistic job cost?
  8. System One model vs LLM: how big is the cost gap?
  9. What these prices leave out
  10. FAQ
  11. Sources

System One models charge only for the text you send them. Output is free on every hosted one we measured, and input list prices run from $0.019 to $0.10 per million tokens. On DecideBench v2, our 450-decision test of 9 November 2026, a million decisions on the hosted ones cost $29.82 to $140.01. Claude Opus 4.8, a general chat model, cost $26,542.70.

If you want the most right answers per dollar, pick Drex Max: 94 of every 100 right at $91.00 per million decisions. If 82 in 100 is good enough, s1-fast does it for $29.82. If the data must stay on your own hardware, OpenJev Flash 9B gets 75 in 100 for $41.33 of GPU time.

  • What you pay for: input tokens only. A token is a chunk of text, about three quarters of a word.
  • Cheapest hosted list price: s1-fast at $0.019 per million input tokens (Decision Models).
  • Dearest hosted list price: GPT-6 Luna at $0.10 per million input tokens (OpenAI).
  • The gap to a chat model: Claude Opus 4.8 cost 291.7 times as much as Drex Max per decision. It got about 5 more right in every 100.
  • The biggest line on the bill: on GPT-6 Luna, the worked examples sent with each question were 61% of it.

A System One model, or decision model, reads a piece of text, a question and a short list of options. It returns one option, with a probability for each. TypeSafe's Jev made the category popular in September 2026. What is a System One model covers how they work. This post is about what they cost. The full rankings are in the best System One models of 2026.

What do System One models charge per token?

Every hosted entry bills input only, at a few cents per million tokens. Here are the list prices DecideBench v2 used, with the exact model served, the vendor page and the date we recorded each price.

EntryModel served$ per M input$ per M outputPrice sourceAs of
s1-fast (Decision Models)s1-fast-20.0190Decision Models pricing2026-10-25
Jev (TypeSafe, via AI Space)typesafe-jev-1.13.00.0420TypeSafe launch post2026-09-25
Drex 2.0 (Nace.AI)drex-v2.00.060Nace.AI pricing reference2026-10-14
Drex Max (Nace.AI)drex-max-v1.10.060Nace.AI Drex Max guide2026-10-19
REAX 1 (Decision Models)reax-1-202610210.080Decision Models pricing2026-10-25
s1-pro (Decision Models)s1-pro-30.080Decision Models pricing2026-10-25
GPT-6 Luna (OpenAI Decisions)gpt-6-luna0.100OpenAI Decisions guide2026-10-08
Claude Opus 4.8 (reference LLM)claude-opus-4-85.0025.00Anthropic pricing2026-10-06

List prices change, and ours are as of the dates in the last column. We rechecked the vendor pages on 9 November 2026.

Decision Models still lists s1-fast at $0.019 on its Peer-to-Peer tier, and s1-pro and REAX 1 at $0.080. Its EU tier, hosted in the EU, costs slightly more: $0.021 and $0.084.

OpenAI now calls the Decisions API generally available. Uncached input is still $0.10 per million tokens, and "output tokens are not charged". Cached input costs less.

TypeSafe priced Jev at "$0.042 / MTok" when it opened early access on 15 September 2026. We called Jev through AI Space, XY Space's platform, and costed it at that list price.

Nace.AI's docs say output tokens are not billed. The price figures on its pages did not render for us on 9 November, so the Drex prices above are the ones we recorded on the dates shown.

Why is the output free?

A decision model writes no text. It returns one option key and a probability for each option, so there is almost nothing to bill on the way out. TypeSafe's launch post calls Jev's output "too cheap to meter", and Simon Willison put it plainly on 21 September 2026: "Jev charges only for input". Decision Models, Nace.AI and OpenAI price the same way.

A large language model (LLM) is the opposite. Claude Opus 4.8 bills $25 per million output tokens, five times its input price, and with adaptive thinking on, as DecideBench ran it, it writes reasoning before it answers.

What does a million decisions cost?

A million decisions cost $29.82 to $140.01 on the hosted decision models. Self-hosted models that got at least 75 in 100 right cost $41.33 to $386.34. Accuracy here means how many of DecideBench's 450 decisions each model got right; DecideBench is made by XY Space, and so is yev1-4b in the table below.

EntryHostingRight answersCost per M decisionsCost per M correct (derived)
Claude Opus 4.8 (reference)API99.3%$26,542.70$26,729.81
Drex MaxAPI94.4%$91.00$96.40
REAX 1API92.4%$120.29$130.18
pplx-decider v1.2 27Bself-hosted92.2%$386.34$419.02
s1-proAPI91.6%$120.93$132.02
JevAPI86.2%$75.04$87.05
Clef 27Bself-hosted85.3%$216.42$253.72
GPT-6 LunaAPI84.0%$140.01$166.68
s1-fastAPI82.0%$29.82$36.37
Drex 2.0API80.9%$87.07$107.63
yev1-4b (ours)self-hosted79.3%$71.15$89.72
OpenJev Flash 9Bself-hosted75.3%$41.33$54.89
Julia-1 144Mself-hosted26.9%$9.13$33.94

The last column divides the cost by the share answered correctly: Drex Max is $91.00 / 0.944 = $96.40 per million right answers. We computed it; DecideBench does not report it.

Cost per million decisions, hosted System One models
Cost per million decisions, hosted System One models
Cost per million decisions
s1-fast$29.82
Jev$75.04
Drex 2.0$87.07
Drex Max$91
REAX 1$120
s1-pro$121
GPT-6 Luna$140

Source: DecideBench v2, measured 9 November 2026, at each vendor's dated list price

The hosted bills sit close together because the billed tokens do. Divide each cost by its list price and you get the tokens each vendor counted per decision. GPT-6 Luna counted the fewest, 1,400. Jev counted the most, 1,787. The other five fell between 1,451 and 1,570. So the price per token does most of the sorting.

Which model is cheapest at the accuracy you need?

Pick the accuracy your job needs first, then the cheapest model that clears it. Read this way, the table gives a short answer for each bar.

You need at leastCheapest hostedCheapest self-hosted
90 right in 100Drex Max, $91.00 (94.4%)pplx-decider 27B, $386.34 (92.2%)
85 right in 100Jev, $75.04 (86.2%)Clef 27B, $216.42 (85.3%)
80 right in 100s1-fast, $29.82 (82.0%)Clef 27B, $216.42 (85.3%)
75 right in 100s1-fast, $29.82 (82.0%)OpenJev Flash 9B, $41.33 (75.3%)

Three things stand out. Drex Max is cheaper than REAX 1 and s1-pro while getting more right, so above 90% nothing hosted competes with it on price. Jev costs less than Drex Max per decision, $75.04 against $91.00, but gets about 8 fewer right in every 100; Drex Max vs Jev covers where those misses fall. And s1-fast is the floor for anything worth using: Julia-1 costs $9.13 but scores 26.9%, next to random guessing, which scored 25.1% on this test.

The same logic, applied across LLMs and coding models, is in the cheapest model per solved task.

How much of the bill is worked examples?

On the one model where we can split the bill, the examples were most of it. DecideBench sends one solved example per option before every item. The item itself is short: a median of 893 characters of input. The bill counted around 1,500 tokens per decision, and the rest is the question, the options and the examples.

GPT-6 Luna is the one hosted model DecideBench also ran without examples, so it shows the split directly. From the billed tokens in the two run files:

GPT-6 LunaTokens per decisionCost per M decisionsRight answers
With one example per option1,400.1$140.0184.0%
Without examples546.5$54.6583.3%

The examples were 61% of Luna's bill and bought 0.7 points, less than one more right answer in every 100. The cost without examples (zero-shot) is our arithmetic on the recorded tokens at the $0.10 list price; DecideBench's published cost column reports only the run with examples. For every other hosted model the cost without examples is n/a: v2 did not run them that way, and Jev's run without examples was skipped because AI Space had no Jev credits.

What this means for your own bill: test with and without examples on a few hundred of your own items before you commit. GPT-6 Luna barely noticed them. TEV, a self-hosted model also run both ways, dropped from 74.9% to 69.6% right without them.

How is self-hosted pricing worked out?

Self-hosted models have no list price, so DecideBench prices them by GPU time. A run's wall-clock time for all 450 items is charged at $3.36 an hour, the on-demand price of one NVIDIA RTX PRO 6000 Blackwell graphics card on AWS g7e.2xlarge. Each run kept 4 requests going at once.

Worked through for pplx-decider 27B: 450 decisions took 186.1 seconds. That is $0.174 of GPU time, or $386.34 per million. OpenJev Flash 9B finished the same 450 in 19.9 seconds, hence $41.33.

A busier server is cheaper per decision. Each self-hosted entry ran alone on its own GPU. A server kept full with more traffic would cost less per task.

Batching changes the bill. Perplexity ships pplx-decider as a Python library, so DecideBench wraps it in a server that groups queued requests into one call. Unbatched, a million decisions cost $875. Batched, they cost $386.34. Accuracy barely moved: 92.4% unbatched, 92.2% batched. Median time per request fell from 3,443 ms to 1,528 ms.

Latency does not compare across the two. Self-hosted times were measured on the GPU box, with no network in them. Hosted times include the network.

Open source System One models covers which ones you can download and run.

What does a realistic job cost?

Take a support inbox that triages one million emails a month. At list price, the monthly bill is the per-million figure itself, and a year is twelve times that. This is list-price arithmetic on DecideBench's inputs, not a quote; your bill scales with how long your emails are and how many options you offer.

ModelPer monthPer yearEmail triage, right answers
s1-fast$29.82$357.8486%
OpenJev Flash 9B (self-hosted)$41.33$495.9680%
Jev$75.04$900.4888%
Drex Max$91.00$1,092.0098%
GPT-6 Luna$140.01$1,680.1288%
Claude Opus 4.8$26,542.70$318,512.4098%

The last column covers the email-triage questions alone, 50 of the 450, so treat each figure as rough: one item is 2 points. On this family Drex Max matched Opus 4.8 at 98% for 1/292 of the price.

System One model vs LLM: how big is the cost gap?

Claude Opus 4.8 cost $26,542.70 per million decisions, 291.7 times as much as Drex Max. Against Jev the multiple is 353.7, and against s1-fast 890.1. Opus got 99.3% right; Drex Max got 94.4%.

Opus's extra 4.9 right answers in every 100 cost $26,451.70 more per million decisions. That works out at about 54 cents for each extra right answer. Whether that is worth paying depends on what a wrong answer costs you. Where it is, a common pattern is a cascade: the decision model answers first and passes its low-confidence cases to the LLM, which our LLM cascade post covers. Decision model vs LLM compares the two on DecideBench v1, an older and different test.

What these prices leave out

  • Your own inputs. DecideBench's items have a median of 893 characters and 3 to 10 options. Longer text or more options means more tokens and a bigger bill.
  • Untested models. Liquid AI d1-ultra, Microsoft-Decision-1, Upstage Solar Decide and others were not in v2, so we have no cost per decision for them.
  • Discounts. Cached input on OpenAI, existing-customer rates on Decision Models and volume deals are not in these figures.
  • Two names missing from v2. Decision Models retired Quyet-1.0-Large on 3 November 2026. Drex 1.5 was not run separately because Nace.AI now answers requests for it with Drex Max, per the DecideBench changelog.

FAQ

Are System One models free?

No. Every hosted System One model DecideBench measured bills input tokens, from $0.019 to $0.10 per million at the list prices we recorded in September and October 2026. Output is free. Jev opened in early access on 15 September 2026 with a waitlist, but it was priced from launch at $0.042 per million input tokens. Open-weight models are free to download, and you pay for the GPU that runs them.

Why is the output of a System One model free?

A decision model writes no text. It returns one option key and a probability for each option, so the output is a few bytes. TypeSafe, Nace.AI, Decision Models and OpenAI all bill input only. An LLM such as Claude Opus 4.8 bills $25 per million output tokens, five times its input price.

What is the cheapest System One model?

By cost per decision, Julia-1 144M, at $9.13 per million self-hosted. It got 26.9% right, close to random guessing, which scored 25.1%. The cheapest one worth using is s1-fast from Decision Models: 82.0% right at $29.82 per million decisions on DecideBench v2, at $0.019 per million input tokens.

Does Jev cost more than Drex Max?

No. On DecideBench v2, Jev cost $75.04 per million decisions and Drex Max $91.00. Drex Max has the higher token price, $0.06 against $0.042, but Jev counted more tokens per decision. Drex Max got 94.4% right against Jev's 86.2%, so per right answer the two are close: $96.40 for Drex Max, $87.05 for Jev. Jev pricing covers Jev's bill in detail.

How much does a self-hosted System One model cost?

You pay for GPU time, at $3.36 an hour for one RTX PRO 6000 graphics card on AWS. A million decisions cost $41.33 on OpenJev Flash 9B. On pplx-decider 27B they cost $386.34. A busier server costs less per decision, and batching cut pplx-decider from $875 to $386.34.

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…