On this page
- What changed this week?
- Which System One model is the most accurate?
- Which System One model should you pick for your situation?
- What is the cheapest System One model worth using?
- What is the best System One model you can run yourself?
- Which System One model is the fastest?
- Is Jev still the best System One model?
- Where do all System One models fall short?
- Which System One models did we not test?
- The full ranking: 26 decision models on DecideBench v2
- How we tested
- Where XY Space fits
- FAQ
- Sources
The best System One model in 2026, the kind of AI that picks one answer from a list, is Nace.AI's Drex Max. On DecideBench v2, our 450-decision test of 9 November 2026, it got 94 of every 100 right. That is the best of 26, at $91 per million decisions. None matches Claude Opus 4.8, the general AI we use as a yardstick (99.3%, about $26,543).
A System One model reads some text, a question and a short list of options, then returns one option. "Approve this refund or send it to a person" is a typical question. What is a System One model explains the category from scratch.
Want the most accurate hosted API? Drex Max, 94.4% at $91 per million.
Want the cheapest hosted model that still gets 4 in 5 right? s1-fast from Decision Models, 82.0% at $29.82.
Need the model on your own hardware? Perplexity's pplx-decider v1.2 27B, 92.2%.
Need answers fast? OpenAI's GPT-6 Luna, 196 ms, the quickest hosted model.
Already on Jev? It scored 86.2%, fifth of 26. Drex Max is 8.2 points ahead for $15.96 more per million.
- 26 decision models: 7 hosted APIs and 19 open models we ran on one graphics card.
- Every price is per million decisions, so hosted and self-hosted models compare directly.
- DecideBench is ours (XY Space), and so are two of the models in it, yev0-4b and yev1-4b. The items, answers and code are public.
What changed this week?
DecideBench v2 went out on 9 November 2026 with a new, harder test, and every number on this page comes from it. The earlier v1 test (400 short decisions) had most models above 90%, so it could no longer tell them apart. v1 and v2 scores are different tests and do not compare.
Each v2 item looks like what a decision model sees in production. That might be a JSON payload from an AI agent, a raw email with its headers, or a release controller's metrics. Inputs are about three times longer than in v1, and each item has 3 to 10 options.
Three entries changed for other reasons:
- Quyet-1.0-Large is gone. Decision Models retired it on 3 November 2026, and its API now answers HTTP 410.
- Drex 1.5 is no longer separate. Nace.AI now serves Drex Max when you ask for Drex 1.5. All 450 answers matched, so it gets no separate score.
- Clef runs on our own graphics card. Cloudflare publishes Clef's weights, so we ran the model ourselves. Through AI Space, our model gateway, Clef 27B scored 68.4% on the same items, against 85.3% when we ran Cloudflare's weights directly. We report the 85.3%.
Which System One model is the most accurate?
Drex Max is the most accurate decision model we have measured: it got 94 of every 100 decisions right. Three models sit two to three points behind it.
| Model | Company | Right answers | Pairs both right | Cost per million | Median wait |
|---|---|---|---|---|---|
| Claude Opus 4.8 (reference, general AI model) | Anthropic | 99.3% | 98.7% | $26,542.70 | 1,336 ms |
| Drex Max | Nace.AI | 94.4% | 89.3% | $91.00 | 258 ms |
| REAX 1 27B | Decision Models | 92.4% | 87.1% | $120.29 | 545 ms |
| pplx-decider v1.2 27B (self-hosted) | Perplexity | 92.2% | 86.2% | $386.34 | 1,528 ms |
| s1-pro 26B | Decision Models | 91.6% | 85.3% | $120.93 | 418 ms |
"Pairs both right" is the stricter score. Every item has a twin that differs by one small, realistic edit that changes the right answer, such as a file attached to the chat instead of a database table. A model that answers from topic words alone gets one twin wrong, so pair accuracy falls faster than plain accuracy.
| Right answers | Both halves of a pair right | |
|---|---|---|
| Drex Max | 94.4% | 89.3% |
| REAX 1 27B | 92.4% | 87.1% |
| pplx-decider v1.2 27B | 92.2% | 86.2% |
| s1-pro 26B | 91.6% | 85.3% |
| Jev | 86.2% | 72.9% |
| Clef 27B | 85.3% | 72.4% |
| GPT-6 Luna | 84% | 72% |
| s1-fast | 82% | 68.9% |
| Drex 2.0 | 80.9% | 65.8% |
| yev1-4b | 79.3% | 62.7% |
Source: DecideBench v2, measured 9 November 2026
Drex Max is 4.9 points behind Opus 4.8 at about 1/290 of its cost. On 450 items, one point is about 4.5 answers. Treat REAX 1, pplx-decider and s1-pro as a cluster, not a strict order.
Drex Max got every item right in three of nine jobs: approving an AI agent's proposed action, moderating a post and reading a customer's intent. Its weakest job was gating a staged software release, at 74%. Drex Max vs Jev goes through the two side by side.
Which System One model should you pick for your situation?
Pick by the constraint you cannot move: accuracy, price, where the model runs, or speed. Each row below is the best measured answer to one of those.
| If you want | Pick | Right answers | Pairs both right | Cost per million | Median wait |
|---|---|---|---|---|---|
| The most accurate hosted API | Drex Max (Nace.AI) | 94.4% | 89.3% | $91.00 | 258 ms |
| The cheapest hosted model above 80% | s1-fast (Decision Models) | 82.0% | 68.9% | $29.82 | 238 ms |
| The most accurate model you run yourself | pplx-decider v1.2 27B (Perplexity) | 92.2% | 86.2% | $386.34 | 1,528 ms, no network |
| The cheapest self-hosted model above 75% | OpenJev Flash 9B (OpenJev) | 75.3% | 55.6% | $41.33 | 174 ms, no network |
| The fastest hosted API | GPT-6 Luna (OpenAI Decisions) | 84.0% | 72.0% | $140.01 | 196 ms |
| The original | Jev (TypeSafe) | 86.2% | 72.9% | $75.04 | 334 ms |
Self-hosted costs are graphics-card time, and self-hosted wait times were measured on the card with no network in them. Hosted wait times include the network from one client machine. The two columns are close, not identical, comparisons.
Moving off Jev specifically? Best Jev alternatives ranks the same models by what each trades away from Jev. This page ranks the whole category, Jev included.
What is the cheapest System One model worth using?
s1-fast from Decision Models is the cheapest hosted model that still gets 4 in 5 right. It scored 82.0% at $29.82 per million, about a third of Drex Max's price.
s1-fast is on the cost frontier, meaning no other model is both cheaper and more accurate. The other hosted models on that frontier are Jev ($75.04, 86.2%) and Drex Max ($91.00, 94.4%). Every other hosted model is beaten on both counts by one of those three.
s1-fast trails Drex Max by 12.4 points, and the gap sits in two jobs. On triaging a bug report it scored 66% to Drex Max's 94%. On gating a staged release, 46% to 74%. On reading a customer's intent the two are within two points.
Decision Models lists s1-fast at $0.019 per million input tokens, with output free. Per-token prices hide how many tokens each model bills for the same decision, which is why we quote cost per decision. System One model pricing covers every model's list price and what it means per decision.
Nothing cheaper than s1-fast is worth using yet. Julia-1 costs $9.13 per million and Laya $13.96, but both score close to random guessing.
What is the best System One model you can run yourself?
Perplexity's pplx-decider v1.2 27B is the most accurate model you can download and run: 92.2%, level with the second-best hosted API. You pay for it in graphics-card time: $386.34 per million decisions, about 4.2 times Drex Max's price.
We ran every self-hosted model on one NVIDIA RTX PRO 6000 Blackwell graphics card, rented from AWS at $3.36 an hour. Each model ran alone, with 4 requests in flight. A busier server that keeps the card full would cost less per decision. Perplexity ships pplx-decider as a Python library, so we wrote a small server that batches queued requests into one call. Without batching it scored 92.4% at $875 per million and 3,443 ms.
The weights are Apache 2.0 on Hugging Face, and Perplexity's model card says the weights take about 52 GiB of graphics-card memory. What is pplx-decider covers the model and its smaller Flash sibling.
Cheaper self-hosted options, in order of accuracy:
| Model | Company | Right answers | Pairs both right | Cost per million | Median wait (no network) |
|---|---|---|---|---|---|
| Clef 27B | Cloudflare | 85.3% | 72.4% | $216.42 | 928 ms |
| yev1-4b (ours) | XY Space | 79.3% | 62.7% | $71.15 | 275 ms |
| pplx-decider-flash v1.2 4B | Perplexity | 78.2% | 62.2% | $125.34 | 284 ms |
| OpenJev Flash 9B | OpenJev | 75.3% | 55.6% | $41.33 | 174 ms |
We built yev1-4b and we built the test, so weigh its row knowing that. The reason to self-host is usually where the data goes, not price: Drex Max is more accurate than every self-hosted model, and cheaper than pplx-decider and Clef 27B. Open source System One models covers the self-hosted field in depth.
Which System One model is the fastest?
GPT-6 Luna, behind OpenAI's Decisions API, is the fastest hosted model: half of its answers came back within 196 ms. s1-fast (238 ms) and Drex Max (258 ms) follow.
Hosted times include the network from our client machine. Jev was called through AI Space, our gateway, so its 334 ms includes one extra hop. Self-hosted times were measured on the graphics card itself: OpenJev Flash 9B answered in 174 ms there, but you would add your own network on top. Laya (58 ms) and Julia-1 (39 ms) are faster still and near chance, so speed alone is a poor reason to pick them.
Most models saw one solved example per answer option before each question. Luna also ran without those examples and scored 83.3%, against 84.0% with them. That matters for cost: examples are input you pay for on every call.
Is Jev still the best System One model?
No. Jev scored 86.2% and placed fifth of 26, behind Drex Max, REAX 1, pplx-decider and s1-pro. It is still good value: no hosted model is both cheaper and more accurate.
The bigger gap is in the pairs. Jev got both halves of 72.9% of pairs right, against Drex Max's 89.3%. That 16.4-point gap means Jev more often gives the same answer to two items that need different ones. Its weakest jobs were gating a staged release (60%), ruling on a return (76%) and triaging a bug report (78%).
TypeSafe introduced Jev as the first System One model on 15 September 2026, and what is Jev covers how it works. For the price detail, read Jev pricing. We called Jev through AI Space, which served version 1.13.0.
Where do all System One models fall short?
Every decision model falls short on rules that need a number worked out first. Gating a staged software release means computing a crash-free rate or error ratio and checking it against an ordered list of rules. Opus 4.8 got all 50 of those right. No decision model got more than 74%.
| Job | Opus 4.8 | Drex Max | REAX 1 | pplx-decider | Jev | GPT-6 Luna |
|---|---|---|---|---|---|---|
| Gating a staged software release | 100% | 74% | 66% | 64% | 60% | 46% |
| Ruling on a return under a written policy | 100% | 90% | 84% | 86% | 76% | 78% |
| Triaging a bug report | 100% | 94% | 92% | 90% | 78% | 78% |
| Reading a customer's intent | 100% | 100% | 100% | 100% | 96% | 98% |
Ruling on a return, where dates and amounts meet a written policy, is the next widest gap. Reading a customer's intent, by contrast, is close to solved: the top four decision models all scored 100%.
One approach worth testing, which DecideBench did not measure: compute the number in code first and pass the result in the input, so the model only has to read it. If a decision depends on arithmetic you cannot move into code, a general LLM is the safer pick for that step. Decision model vs LLM covers that split.
Each family has 50 items, so a single family's score moves 2 points per answer and carries wide error bars. The overall ranking is firmer than any one row of this table.
Which System One models did we not test?
Several decision models are on the market that DecideBench has no score for. We list them so a missing name is not read as a low score:
- Liquid AI d1-ultra
- Microsoft-Decision-1
- Strands Decider 2B
- Upstage Solar Decide
- meraGPT Decider 1
- Quyet-1.0-Large (retired by Decision Models on 3 November 2026)
- Drex 1.5 (Nace.AI now serves Drex Max under that name)
- The Jeff models (the pinned upstream code they need is gone)
We also did not test general LLMs other than the reference in v2. Any score you see for these models comes from their makers or from another test, and is not ours.
The full ranking: 26 decision models on DecideBench v2
All 26 decision models, plus the reference and a random baseline, measured on 9 November 2026. A dash in Company means an individual published the model.
| # | Model | Company | Hosting | Right answers | Pairs both right | Cost per million | Median wait |
|---|---|---|---|---|---|---|---|
| – | Claude Opus 4.8 (reference) | Anthropic | API | 99.3% | 98.7% | $26,542.70 | 1,336 ms |
| 1 | Drex Max | Nace.AI | API | 94.4% | 89.3% | $91.00 | 258 ms |
| 2 | REAX 1 27B | Decision Models | API | 92.4% | 87.1% | $120.29 | 545 ms |
| 3 | pplx-decider v1.2 27B | Perplexity | Self-hosted | 92.2% | 86.2% | $386.34 | 1,528 ms |
| 4 | s1-pro 26B | Decision Models | API | 91.6% | 85.3% | $120.93 | 418 ms |
| 5 | Jev | TypeSafe | API | 86.2% | 72.9% | $75.04 | 334 ms |
| 6 | Clef 27B | Cloudflare | Self-hosted | 85.3% | 72.4% | $216.42 | 928 ms |
| 7 | GPT-6 Luna | OpenAI | API | 84.0% | 72.0% | $140.01 | 196 ms |
| 8 | s1-fast 26B-A4B | Decision Models | API | 82.0% | 68.9% | $29.82 | 238 ms |
| 9 | Drex 2.0 | Nace.AI | API | 80.9% | 65.8% | $87.07 | 264 ms |
| 10 | yev1-4b (ours) | XY Space | Self-hosted | 79.3% | 62.7% | $71.15 | 275 ms |
| 11 | pplx-decider-flash v1.2 4B | Perplexity | Self-hosted | 78.2% | 62.2% | $125.34 | 284 ms |
| 12 | yev0-4b (ours) | XY Space | Self-hosted | 78.2% | 60.0% | $162.63 | 408 ms |
| 13 | Bespoke-Nimble-9B | Bespoke Labs | Self-hosted | 75.6% | 56.0% | $135.67 | 507 ms |
| 14 | OpenJev Flash 9B | OpenJev | Self-hosted | 75.3% | 55.6% | $41.33 | 174 ms |
| 15 | Tev | Together AI | Self-hosted | 74.9% | 54.7% | $81.88 | 306 ms |
| 16 | imajev-4b | – | Self-hosted | 74.0% | 53.8% | $57.59 | 248 ms |
| 17 | JevK5 v0.3 | – | Self-hosted | 68.0% | 44.0% | $73.02 | 310 ms |
| 18 | Clef-Flash 9B | Cloudflare | Self-hosted | 66.9% | 45.8% | $65.42 | 273 ms |
| 19 | Decider-4B | – | Self-hosted | 63.6% | 42.2% | $63.18 | 254 ms |
| 20 | Kev-4B | – | Self-hosted | 41.8% | 17.8% | $100.81 | 374 ms |
| 21 | GLiNER2.5-Decide 340M | Fastino | Self-hosted | 40.7% | 7.6% | $94.79 | 432 ms |
| 22 | Laya typed-decisions 421M | – | Self-hosted | 39.1% | 10.7% | $13.96 | 58 ms |
| 23 | Kev-9B | – | Self-hosted | 35.8% | 15.1% | $123.46 | 466 ms |
| 24 | Decider-2B | – | Self-hosted | 31.8% | 10.2% | $28.19 | 115 ms |
| 25 | Julia-1 144M | Supersonic Labs | Self-hosted | 26.9% | 2.2% | $9.13 | 39 ms |
| 26 | CLM-v0.1-8B | – | Self-hosted | 26.4% | 1.8% | $16.65 | 67 ms |
| – | Random guessing | – | – | 25.1% | 4.4% | $0.00 | 0 ms |
Self-hosted wait times have no network in them. The four small encoder models (GLiNER2.5-Decide, Laya, Julia-1 and CLM) score 26% to 41%, against 25.1% for random guessing; Laya, Julia-1 and CLM ran without worked examples because they have no place to take them.
How we tested
Every model saw the same 450 decisions in 225 pairs, 50 from each of nine jobs: approving an AI agent's proposed action, routing an agent to a tool, checking a claim against its sources, moderating a post, ruling on a return, gating a staged software release, reading a customer's intent, triaging an inbound email and triaging a bug report. Where a decision depends on a written policy, the question carries the full policy.
Models that take worked examples saw one solved example per option before each item. Hosted costs are the tokens each API billed, at its dated list price. Self-hosted costs are the run's wall-clock time on the graphics card at its hourly price. Median wait is the time per request with 4 in flight.
Claude wrote every item, and Claude Opus, Sonnet and Haiku checked each answer before the runs. That is a real limit: a careful reader could dispute a few labels, and one of Opus 4.8's three misses is arguable. The items are public, so a model trained on them later could score higher; every line carries a canary string asking trainers to exclude it. System One model benchmark explains the method in full.
Where XY Space fits
AI Space, our platform, hosts Jev and open decision models on Western infrastructure. XY Space builds agentic workflows that use decision models for steps like routing and triage, and DecideBench is how we choose which one goes where.
FAQ
What is a System One model?
A System One model is a decision model: it reads an input, a question and a fixed list of options, and returns one option with a score instead of writing text. TypeSafe introduced the name when it launched Jev on 15 September 2026. Because it writes nothing, it bills mainly for input and answers in a few hundred milliseconds. What is a System One model covers it in full.
What other decision models are there besides Jev and GPT-6?
Plenty. On DecideBench v2 the most accurate is Nace.AI's Drex Max (94.4%), followed by Decision Models' REAX 1 (92.4%), Perplexity's open pplx-decider (92.2%) and Decision Models' s1-pro (91.6%). All four beat Jev (86.2%) and OpenAI's GPT-6 Luna (84.0%). Cloudflare's Clef, Nace.AI's Drex 2.0, Together AI's Tev, Bespoke-Nimble and our own Yev models follow. The full table above lists all 26.
Are System One models cheaper than LLMs?
Yes, by a wide margin on this test. Drex Max costs $91 per million decisions, about 1/290 of Claude Opus 4.8's $26,542.70, and s1-fast costs $29.82. The trade is accuracy: Opus scored 99.3%, and no decision model passed 94.4%. DecideBench v2 did not test cheaper general LLMs, so the gap to a small chat model is not measured here.
Can I run a System One model locally?
Yes. 19 of the 26 models we tested are open weights, and we ran each of them on one NVIDIA RTX PRO 6000 graphics card. The most accurate, pplx-decider v1.2 27B, needs about 52 GiB of graphics-card memory for its weights. Smaller 4B models such as yev1-4b and pplx-decider-flash run at 275 to 284 ms on that card, before any network.
Sources
- XY Space, DecideBench repository and v2 results, v2.0, measured 9 November 2026
- XY Space, DecideBench dataset, v2 config, 9 November 2026
- XY Space, DecideBench leaderboard, accessed 9 November 2026
- XY Space, DecideBench changelog, 9 November 2026
- TypeSafe, Introducing System One Models & Jev, 15 September 2026
- Simon Willison, Jev introduces a new shape of LLM: System One, aka Decision Models, 21 September 2026
- Nace.AI, Drex Max guide, accessed 9 November 2026
- Decision Models, Pricing, accessed 9 November 2026
- OpenAI, Decisions API guide, accessed 9 November 2026
- Perplexity, pplx-decider-v1.2 model card, accessed 9 November 2026
- AWS, Amazon EC2 G7e instances, accessed 9 November 2026
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 YongShare 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