67.59 Score: Nace AI's Drex Max Outperforms Jev With $0.06/M Input Tokens For Computer Agents

Nace AI has released Drex Max, a multimodal decision model designed for agents that interact with computers and browsers. According to MarkTechPost, the model reads text and screenshots to return calibrated probabilities rather than generating new text. This architecture allows agent builders to make one cheap, typed call per click, streamlining the decision-making process for automated systems.

The risk

The primary consideration for developers is the model's specific performance profile. Nace reports that Drex Max achieves 67.59 on the public Decision Index 0.3.2, which is the top score listed for any decision model in their family. It also posts 98.2 on BFCL tool calling benchmarks. However, the model shows lower scores on other metrics, including 19.0 on HLE and 72.5 on MMLU-Pro, compared to 82.1 for Jev 1.14.

View the original post on X

Who is exposed

This release targets developers building computer-use agents. Drex Max is the only model in the Drex family that reads images, making it relevant for applications requiring visual context. It is available only via a hosted API on the Nace Console. There are no open weights or quantized builds announced, meaning users must rely on the vendor's infrastructure. The model supports up to 262,144 tokens per question on the API, including image tokens.

Cost and compliance

Nace lists the pricing for Drex Max at $0.06 per million input tokens. Each image in a request costs up to 1,602 tokens in every question. The model handles up to eight images per request, with a limit of 16 MiB each and 32 MiB total. All images are resized to at most 1,638,400 pixels. The request body is limited to 4 MiB, and users can submit up to 16 questions per request.

What should you do?

Developers should evaluate whether the model’s strengths in tool calling and decision indexing align with their specific agent requirements. Teams should review the lower scores on HLE and MMLU-Pro to determine if those limitations impact their use cases. Users must integrate via the Nace Console API, as no local deployment options are currently available.

The background in plain language

Decision models differ from standard language models because they do not generate free-form text. Instead, they answer specific, typed questions about a given state. A user sends a state and named questions, which can be yes/no, choice, or score types. The model returns a probability distribution for each answer in a single call. This approach reduces ambiguity and computational overhead, as the model only selects from predefined options provided by the developer.

What to do and how to stay safe: Nace AI

• Review the model's benchmark scores, particularly the lower HLE and MMLU-Pro results, to ensure they fit your application's needs.
• Calculate the token cost implications of sending multiple images, as each image adds significant token overhead to every question.
• Test the hosted API integration to verify that the 4 MiB request body limit accommodates your typical workload size.
• Monitor the vendor's documentation for any updates on open weights or quantized builds if local deployment is required.

Step-by-step guide: AI Governance Explained: Building Rules for Autonomous Systems

General security guidance from the Vector Tribune 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 write text; it returns calibrated probability distributions for predefined answer options.

How many images can Drex Max process in a single request?

The model supports up to eight images per request, with a maximum size of 16 MiB each and 32 MiB total.

Is Drex Max available as open weights?

No, Nace AI has not announced any open weights or quantized builds; the model is available only via a hosted API.