Key points
- Drex Max achieves a 67.59 score on Decision Index 0.3.2, outperforming TypeSafe’s Jev 1.14 by 6.9 points.
- The model reads text and images, returning probability distributions for predefined question types without generating new text.
- Pricing is set at $0.06 per million input tokens, with support for up to 262,144 tokens per API request.
Nace AI has released Drex Max, a multimodal decision model designed specifically for computer-use agents that need to interpret screen content and textual data. According to MarkTechPost, this new model reads both text and screenshots to return calibrated probabilities, marking a shift from generative models to specialized decision-making tools. The release aims to provide agent builders with a cost-effective method for handling complex interface interactions through typed calls rather than open-ended text generation.
How it unfolded
- Nace AI announced Drex Max as the most accurate model in its Drex family and the only one capable of reading images.
- The company reported a score of 67.59 on the public Decision Index 0.3.2, positioning it as the top-scoring decision model in its lineup.
- Performance benchmarks showed Drex Max scoring 92.2 on JevBench, significantly higher than the 86.7 achieved by TypeSafe’s Jev 1.14.
Who is affected
Developers building automated agents that interact with computers and browsers are the primary audience for this release. According to the source material, these builders benefit from a single cheap, typed call per click, reducing computational overhead. The model is available exclusively via a hosted API on the Nace Console, identified by the pinned ID "drex-max". No open weights or quantized builds have been announced, meaning users must rely on Nace’s infrastructure for access and updates.
View the original post on X
What happens next
Users can integrate Drex Max using the existing POST /v1/systemone endpoint by specifying the model name. The system allows up to eight images per request, with each image resized to a maximum of 1,638,400 pixels. Each image contributes up to 1,602 tokens to the cost calculation for every question processed. With a maximum request body of 4 MiB and support for up to 16 questions per request, the model is structured for high-throughput agent workflows.
The background in plain language
Decision models differ from generative AI by not creating new text. Instead, they analyze provided data to select from predefined options. Teams usually configure these systems with specific question types, such as yes/no queries or labeled choices. The model returns a probability distribution for each option, allowing software to make deterministic choices based on visual and textual inputs. This approach reduces ambiguity in automated workflows.
What to do and how to stay safe: Nace AI
- Review current agent architectures to identify tasks that rely on open-ended text generation for interface navigation.
- Evaluate whether switching to a decision-based model could reduce token costs and improve response consistency for click-based actions.
- Test the hosted API with sample screenshots to verify that image resizing and token counting align with your budget constraints.
- Monitor the pinned model ID to ensure routing remains stable, as updates may not automatically apply to latest versions.
Step-by-step guide: Injection attacks examples: how prompt leaks break AI systems
General security guidance from the Circuit Observer newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is the cost of using Drex Max?
Nace AI reports the cost as $0.06 per million input tokens.
Can Drex Max generate new text responses?
No, the model does not generate tokens; it returns probability distributions for predefined answer options.
How does Drex Max handle images?
It accepts up to eight images per request, resizing them to a maximum of 1,280 x 1,280 pixels.
