Drex Max.
The decision model that sees.
Screens, images and video in. A calibrated probability for every option out. Built for agents that use computers and browsers.
- #1
- of 17 decision models on Artificial Analysis
- 118×
- faster than Claude Haiku 5.5 (Max)
| Model | Creator | Type | Decision Index | Cost per 1M Tasks (USD) |
|---|---|---|---|---|
| Claude Haiku 5.5 (Max) | Anthropic | Language model | 77.1 | $6.2k |
| GPT-6 Luna (Max) | OpenAI | Language model | 76.4 | $1.5k |
| Drex Max v1.1 | Nace.AI | Decision model | 73.4 | $102 |
| GLiDE v1.2 | Fastino | Decision model | 72.8 | $780 |
| Mammoth 27B | Base Labs | Decision model | 72.4 | $262 |
| Decider v1.2 | Perplexity | Decision model | 70.5 | $68 |
| d1-ultra | Liquid AI | Decision model | 70.4 | $132 |
| GPT-6 Luna (Low) | OpenAI | Language model | 69.2 | $484 |
| SubQ Snap 1.1 | Subquadratic | Decision model | 69.2 | $136 |
| Elephant 9B | Base Labs | Decision model | 67.5 | $92 |
| Jev 1.14 | TypeSafe | Decision model | 67.0 | $156 |
| Claude Haiku 5.5 (Low) | Anthropic | Language model | 66.7 | $858 |
| Celeris-1.1 Decision | Celeris | Decision model | 63.5 | $14 |
| Drex 2.0 | Nace.AI | Decision model | 63.4 | $199 |
| Mercury Decide | Inception | Decision model | 63.0 | $69 |
| Microsoft-Decision-1 | Microsoft | Decision model | 62.8 | $142 |
| GPT-6 Luna (Decisions) | OpenAI | Decision model | 61.7 | $337 |
| d1 | Liquid AI | Decision model | 61.6 | $63 |
| Decider Flash v1.2 | Perplexity | Decision model | 61.6 | $196 |
| Clef | Cloudflare | Decision model | 60.2 | $859 |
| Clef-Omni | Cloudflare | Decision model | 59.3 | $527 |
The Drex family. Each one built for a job.
Same request, same typed answers, one API key. Pick the model by what your agent looks at and where it runs.
- Screens
- Images
- Charts
- Video
NewMaxDrex Max v1.1
drex-maxThe frontier. Decides from screens, images and video.
Reads the screenshot and picks the next click. The most accurate Drex.
2.0Drex 2.0
drex-v2.0Open weights. Run it on your own hardware.
Download the weights and serve them on a single accelerator.
1.5Drex 1.5
drex-v1.5Lean and quick, for high-volume text routing.
Routing, triage and classification where every millisecond counts.
Ahead on the Public Decision Index. Drex Max scores 67.5, +6.8 over Jev 1.14.
The Public Decision Index, a community leaderboard on Hugging Face, scores tests of deciding rather than writing, chance-corrected so a random guess scores zero. Frontier chat models are shown for reference: Drex Max answers in a median 172 ms for $0.03 per thousand decisions.
| Decision models | Frontier LLMs, for reference | ||||||
|---|---|---|---|---|---|---|---|
| Drex Max v1.1 | Drex 2.0 | Drex 1.5 | Jev 1.14 | GPT-6 Luna Decisions | GPT-6 Sol | Claude Opus 5.5 | |
| Decision | |||||||
| Public Decision Indexpublic | 67.5 | 61.3 | 58.1 | 60.7 | 47.8 | – | – |
| JevBench | 93.9 | 85.7 | 84.8 | 86.7 | 91.0 | 99.1 | 100.0 |
| Decision Bench | 94.2 | 88.3 | 86.8 | 92.3 | 90.5 | 94.7 | 95.6 |
| DecideBenchitem | 99.0 | 96.5 | 90.0 | 98.0 | 97.2 | 99.2 | 98.8 |
| DecideBenchpair | 98.0 | 93.5 | 80.5 | 96.0 | 94.5 | 98.5 | 98.5 |
| Game Benchwin % | 65.2 | 60.3 | 58.7 | 50.0 | 57.5 | 65.0 | 63.0 |
| Knowledge | |||||||
| MMLU-Pro | 75.2 | 72.8 | 58.6 | 82.1 | 58.1 | 93.0 | 94.0 |
| HLE | 18.2 | 12.8 | 11.4 | 19.4 | 12.3 | 39.0 | 56.0 |
| CRUXEval | 86.8 | 84.6 | 76.7 | 74.7 | 66.4 | 100.0 | 95.0 |
| Reasoning | |||||||
| ARC-Challenge | 96.8 | 96.8 | 95.9 | 97.6 | 96.9 | 99.0 | 98.0 |
| ARC-Easy | 99.2 | 98.6 | 97.9 | 99.3 | 99.0 | 100.0 | 100.0 |
| HellaSwag | 98.1 | 95.9 | 92.1 | 94.3 | 91.4 | 93.0 | 96.0 |
| Agentic | |||||||
| OSWorld-Verifiedautomation | 0.787 | 0.698 | 0.676 | 0.734 | 0.747 | 0.756 | 0.790 |
| OSWorld-V2automation | 0.729 | 0.636 | 0.640 | 0.687 | 0.659 | 0.744 | 0.761 |
| BFCLtool calling | 97.8 | 96.6 | 94.9 | 96.4 | 94.2 | 94.7 | 97.3 |
| ToolRettool retrieval | 71.1 | 67.3 | 70.2 | 65.0 | 60.4 | 72.7 | 74.8 |
| When2Call | 91.8 | 86.5 | 89.7 | 81.7 | 70.2 | 78.3 | 86.7 |
| RAGTruth | 85.8 | 82.5 | 82.7 | 75.1 | 77.0 | 72.5 | 69.8 |
| Efficiency | |||||||
| Median API latency | 172 ms | 151 ms | 145 ms | 196 ms | 326 ms | ~2.0 s | ~3.5 s |
| Costper 1k decisions | $0.03 | $0.06 | $0.04 | $0.042 | $0.10 | $3.33 | $10.17 |
Bold marks the best decision model in each row. GPT-6 Sol and Claude Opus 5.5 are general-purpose frontier models, shown for reference; they are not scored on the Public Decision Index. Its score is the public score on edition 0.4.1, where Drex Max is listed as Drex Max v1. Latency is the median API time per decision, as the caller sees it, network included; cost is per 1,000 decisions, with Drex Max v1.1 at its 50% launch price.
Join the Drex communityBuilt for screens. Computer use starts with reading the screen right.
A computer-use agent makes dozens of decisions per task, each one from a screenshot. Drex Max makes each in one pass, from screens, charts, forms, photos and video.
Perception
- Drex Max v1.1
- 86.3
- GPT-6 Luna Decisions
- 73.8
- Drex Max v1.1
- 75.0
- GPT-6 Luna Decisions
- 63.7
- Drex Max v1.1
- 83.5
- GPT-6 Luna Decisions
- 75.0
- Drex Max v1.1
- 93.4
- GPT-6 Luna Decisions
- 83.8
Documents & charts
- Drex Max v1.1
- 86.6
- GPT-6 Luna Decisions
- 78.8
- Drex Max v1.1
- 98.4
- GPT-6 Luna Decisions
- 93.8
- Drex Max v1.1
- 95.9
- GPT-6 Luna Decisions
- 97.5
Agentic
- Drex Max v1.1
- 89.2
- GPT-6 Luna Decisions
- 81.2
Reasoning & safety
- Drex Max v1.1
- 58.6
- GPT-6 Luna Decisions
- 32.5
- Drex Max v1.1
- 53.8
- GPT-6 Luna Decisions
- 41.2
- Drex Max v1.1
- 72.8
- GPT-6 Luna Decisions
- 71.2
And the decisions behind the screen
Finance, banking and legal, where a wrong call costs money.
Finance
Banking
LegalVision scores are Drex Max v1.1 against GPT-6 Luna Decisions on the public Decision Index Vision suite, 11 benchmarks. Finance, banking and legal scores are Drex Max v1.1 against Jev 1.14 on each public benchmark. The Public Decision Index above is edition 0.4.1.
Trusted by
- Big Fouraccounting firm
See it answer. Real cases, recorded replies.
Pick a case. The same text and questions went to both columns, and neither wrote a sentence back.
A topic choice and an escalate yes/no on the customer message from Jev's own docs.
Customer: I was charged twice last Tuesday and I am furious. Order #4419. I want a refund today.
topic choice
What is the issue about?
billing · bug · shipping
urgent yes/no
Escalate to a human now?
Drex
72 tokens read
topic choice
urgent yes/no
Jev 1.13.0
365 tokens read
topic choice
urgent yes/no
Run your own case.
Paste a case, name the options, and get the probabilities back. Free to try, no card needed.
On topic, both columns are all but certain; the other questions show how far apart they can be.
- SOC 2 Type II
- Nace undergoes regular Service Organization Controls audits (SOC 2 Type II). Request SOC 2
- Deployment
- Self-hosted or hybrid, so the cases Drex reads never leave your network.
Frequently asked questions
What is Drex, and how is a decision model different from an LLM?
Drex is a decision model: you send the facts of a case and typed questions, and it returns a probability for every option in a single pass. An LLM writes text token by token and names an answer at the end. With no token stream and no JSON to repair, truncation, malformed output and invented options do not arise. Drex can only answer with the options you gave it.
What kinds of questions can Drex answer?
Three types: a choice between named options, a yes/no question, and a score on a scale you define. Each comes back with probabilities: one for every option, the probability of yes, or the expected level with the spread across levels. Several questions about one case can share a single pass.
How does Drex compare to Jev?
On the Artificial Analysis Decision Index v2, as of Nov 9, 2026, Drex Max v1.1 scores 73.4 against 67.0 for Jev 1.14, and leads it on document decisions (AA-Paperclip 72.8 against 57.5), tool calling (When2Call 95.3 against 84.4) and consistency (AA-Symmetry 85.0 against 79.3). It trails Jev 1.14 on MMLU-Pro, HLE, ARC-Challenge and ARC-Easy. The API is wire-compatible with Jev, so switching takes three environment variables.
How big is Drex?
Drex 2.0 has 35B parameters, with about 3B active per token, so it holds the memory of a 35B model and does the work of a much smaller one. It fits on a single accelerator. The bf16 weights take about 70 GB and the 8-bit build about 37 GB, and it reads up to 128K tokens of context in one request.
Is Drex trained on the benchmarks it's measured on?
No. Drex is trained on the official training splits of the benchmarks, alongside verifiable procedural data, and evaluated only on the held-out splits with the official scorer. That is how the leaderboard is designed to be run. It is strongest where it has seen the kind of decision before, which is why tuning it on your decisions pays.
Can Drex run inside our own environment?
Yes. Drex runs on Nace-managed infrastructure, in your own cloud account, on-premises or at the edge. Drex 2.0 ships open weights and fits on a single accelerator, so you can run it on your own hardware. We can also tune Drex on your labelled outcomes and ship you the weights.
Put a decision model where the guesswork is
Free to try today. Tuned on your decisions and served in your environment when you are ready.