Nace AI launches Drex Max decision model for computer-use agents

Nace AI has released Drex Max, a new multimodal decision model designed for computer-use agents, according to a report by MarkTechPost. The system processes text and screenshots to return calibrated probabilities rather than generating natural language text. This release positions the model as the most accurate option in the Drex family for handling visual inputs.

The model achieves a score of 67.59 on the public Decision Index 0.3.2, which Nace cites as its highest benchmark result. This performance places it 6.9 points above TypeSafe’s Jev 1.14. For developers building automated agents, the model provides a single, low-cost API call to determine the next action, reducing computational overhead during execution cycles.

View the original post on X

Technical Specifications and Performance

Drex Max operates through a hosted API only, with no open weights or quantized builds announced by the vendor. The system accepts up to 262,144 tokens per question, including image tokens, via the POST /v1/systemone endpoint. Users must specify the model ID explicitly as "drex-max," as the "drex-latest" identifier does not route to this specific version.

The model handles visual data by accepting an array of up to eight images per request, with a maximum size of 16 MiB each and 32 MiB total. Each image is resized to a maximum of 1,638,400 pixels. Every image adds up to 1,602 tokens to the cost of each question processed. The system supports up to 16 questions per request within a 4 MiB request body limit.

Benchmark Results and Limitations

Nace reports that Drex Max excels in tool-calling and agentic tasks, scoring 98.2 on BFCL and 0.780 on OSWorld-Verified. It also achieves 92.2 on JevBench, compared to 86.7 for Jev 1.14. However, the model shows weaker performance on general knowledge tasks, scoring 72.5 on MMLU-Pro versus 82.1 for the competitor.

The model struggles with HLE benchmarks, achieving a score of only 19.0. Nace states that the model trains on official training splits and is scored on held-out splits. The output format is strictly limited to probability distributions for predefined options, such as yes/no, choice, or score types, without generating free-form text.

What should you do?

Security and technology professionals should evaluate this model if they are building agents that require visual context. Developers must verify that the cost structure of $0.06 per million input tokens aligns with their operational budgets, especially given the high token cost associated with image processing.

What to do and how to stay safe: Nace AI

• Review the API documentation to understand how image tokens are counted and priced in your specific use case.
• Test the model against your specific decision-making workflows to validate the 67.59 Decision Index score in real-world conditions.
• Monitor the model's performance on general knowledge tasks if your agents require broad contextual understanding beyond visual inputs.
• Ensure your application handles the 4 MiB request body limit and image resizing constraints correctly to avoid API errors.

Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Does Drex Max generate text responses?

No, Drex Max does not generate tokens or text; it returns probability distributions for predefined answer types.

How much does it cost to use Drex Max?

Nace AI charges $0.06 per million input tokens for the Drex Max model via its hosted API.

Is Drex Max available as open source?

No, the model is available only through a hosted API with no open weights or quantized builds announced.