# Drex Max: Nace AI’s New Multimodal Decision Model for Computer-Use Agents
## Introduction
Nace AI has introduced Drex Max, a decision-focused artificial intelligence model designed for agents that interact with computers and web browsers. Unlike traditional language models that generate free-form text, Drex Max takes a different approach: it reads text and images, then returns calibrated probabilities for specific answer options provided by the user. It does not write or generate text on its own.
This makes Drex Max particularly well-suited for tasks where an AI agent needs to make quick, structured decisions — such as choosing the next button to click on a screen or selecting the most relevant option from a list. Nace AI reports that Drex Max achieves a score of 67.59 on the public Decision Index 0.3.2, which is 6.9 points ahead of the next best competing model from TypeSafe AI.
## How Drex Max Works
Drex Max operates as a “decision model,” which means it does not produce new text. Instead, it works with the input you give it and returns a probability distribution across the answer options you specify.
Here is a simplified breakdown of the process:
1. **You provide a state** — this can be text or JSON describing the current situation.
2. **You include images** — up to 8 still images in PNG, JPEG, or WebP format per request.
3. **You ask structured questions** — each question must be one of three types:
– **Yes/No** (noul)
– **Pick a label** (choice)
– **Ordered scale** (score)
4. **The model responds** with a probability for each option in a single HTTP call.
This design means Drex Max can only answer with the options you give it. It cannot invent new answers. This makes it highly predictable and easy to integrate into agent pipelines.
## Working with Images
One of Drex Max’s standout features is its ability to process images. Each image is resized to a maximum of 1,280 × 1,280 pixels, and every question in a request sees the same set of images. You can refer to them simply as “Image 1,” “Image 2,” and so on in your text prompts.
Key rules for images:
– Up to 8 images per request
– Each image can be up to 16 MiB, with a 32 MiB total limit
– Each image consumes up to 1,602 tokens in every question
– Supported formats are PNG, JPEG, and WebP only
## Performance and Benchmarks
Nace AI has shared benchmark results across several categories. It is worth noting that these scores are self-reported and based on public evaluation suites.
### Decision Benchmarks
– **Decision Index 0.3.2 (public):** 67.59
– **JevBench:** 92.2 (compared to 86.7 for Jev 1.14)
– **Decision Bench:** 94.3
– **Game Bench win percentage:** 64.0
### Agentic Benchmarks
– **BFCL tool calling:** 98.2
– **OSWorld-Verified:** 0.780
– **When2Call:** 85.2
– **RAGTruth:** 84.8
### Vision Benchmarks
– **Decision Index Vision (public):** 74.5
– **Mind2Web:** 88.6
– **CV-Bench:** 86.3
### Where Drex Max Trails
– **MMLU-Pro:** 72.5 (vs. 82.1 for Jev 1.14)
– **HLE:** 19.0 (vs. 19.4 for Jev 1.14)
– **ARC-Challenge:** 97.4 (vs. 97.6 for Jev 1.14)
– **Hateful Memes:** 70.6 (vs. 71.2 for GPT-6 Luna Decisions)
## Speed and Pricing
Drex Max is designed for speed. Nace AI reports a median latency of 174 milliseconds per screenshot decision. This compares favorably to Jev 1.14, which averages 253 ms on the same set of questions.
Pricing is structured around input tokens only:
– **$0.06 per 1 million input tokens**
– Output tokens are not billed
– Failed requests cost nothing
– New accounts receive $25 in signup credit
## How Drex Max Compares to Other Models
| Feature | Drex Max | Jev 1.14 | Drex 2.0 | d3-flash |
|—|—|—|—|—|
| Developer | Nace AI | TypeSafe AI | Nace AI | vLLM Semantic Router |
| Inputs | Text and images | Text only | Text | Text, images, and video |
| Context Window | 262,144 tokens per question | 64K per request | 16,384 default | Not disclosed |
| Price per 1M input tokens | $0.06 | $0.042 | $0.06 | Free trial (self-hosted) |
| Decision Index (public) | 67.6 | 60.7 | 61.25 | 57.8 |
| Vision Index (public) | 74.5 | Not disclosed | N/A | 69.9 |
| License | Proprietary (hosted) | Proprietary (hosted) | Open RAIL-M | Apache-2.0 |
Drex Max’s main advantages are its multimodal capability, large context window, and strong decision benchmark scores. It is hosted exclusively through Nace’s console and does not have open weights or quantized builds available.
## Key Takeaways
– Drex Max is Nace AI’s first multimodal decision model.
– It processes both text and images and returns calibrated probabilities — not generated text.
– It leads the public Decision Index 0.3.2 with a score of 67.59.
– It is fast at roughly 174 ms per screenshot decision.
– Pricing is affordable at $0.06 per 1M input tokens with no output token charges.
– The model is hosted only; weights are not publicly available.
## Frequently Asked Questions
**Is Drex Max open source?**
No. Drex Max is proprietary and hosted exclusively on Nace AI’s console. There are no open weights or downloadable versions announced.
**What types of questions can Drex Max answer?**
Drex Max supports three question types: yes/no questions, label selection questions, and ordered scale questions. You must provide the options — the model cannot generate new answers.
**How many images can I include in a single request?**
You can include up to 8 images per request, each in PNG, JPEG, or WebP format, with a maximum size of 16 MiB per image and 32 MiB total.
**Does Drex Max generate text?**
No. Drex Max is a decision model that returns probability distributions over the answer options you provide. It does not produce free-form text output.
**How does Drex Max compare in price to Jev 1.14?**
Drex Max costs $0.06 per 1M input tokens, while Jev 1.14 costs $0.042 per 1M input tokens. However, Drex Max offers a much larger context window and multimodal input support.
**What is the context window for Drex Max?**
Drex Max supports up to 262,144 tokens per question, which includes both text tokens and image tokens.
**Are the benchmark scores officially verified?**
The benchmark scores are self-reported by Nace AI and are based on public evaluation splits. They have not been independently verified by a third party.
## Conclusion
Drex Max represents a focused, practical approach to AI decision-making for computer-use agents. By combining multimodal input with structured, probabilistic outputs, it offers a reliable and fast tool for agents that need to navigate screens and make structured choices. Its strong benchmark performance, affordable pricing, and image processing capabilities make it a compelling option for developers building agent-based applications. However, its closed-source nature and specialized design mean it is best suited for teams looking for a hosted decision model rather than an open, general-purpose language model.
Thank you for reading
