Nace AI is expanding its decision-model technology with Drex Max, a multimodal model designed to evaluate text and images and return structured decisions. The release adds visual input to the company’s Drex decision-model approach, opening up new possibilities for developers building computer-use agents and automated workflows.
Unlike a conventional chatbot that generates a sentence or a sequence of instructions, Drex is designed to score predefined answers to a question. According to Nace AI’s documentation, Drex Max accepts text and image inputs and returns probabilities for the available answers. This approach could help AI agents interpret screenshots, assess interface states, and select from a defined set of possible actions.
The development matters as computer-use agents become more capable of interacting with software interfaces. These systems need to identify what is happening on a screen, decide which action is appropriate, and evaluate whether a task can safely proceed. Drex Max offers a decision-focused component that developers can incorporate into those workflows, although it is not, by itself, a complete computer-control agent.
What Is Drex Max?
Drex Max is Nace AI’s multimodal version of its decision model. The official documentation describes it as the most accurate model in the Drex family and confirms that it can read both text and images.
The underlying idea differs from the usual generative AI workflow. A conventional large language model typically produces text token by token, even when the application ultimately needs a simple classification or selection. Developers may then have to parse that output and handle responses that do not match the expected format.
Drex follows another approach. An application supplies a state containing the relevant information and defines questions with accepted answer options. The model evaluates those options and returns probabilities rather than generating a free-form response.
For example, an application processing a customer complaint could ask whether the customer needs a refund, a replacement, or additional information. Drex Max can evaluate the case alongside relevant visual evidence, such as a product photograph, and score the available outcomes.
This does not mean the model always makes the correct decision. Its probabilities are outputs to evaluate, and developers still need appropriate validation, confidence thresholds, and escalation procedures for uncertain or consequential cases.
Why Multimodal Input Matters for Computer Use Agents
Computer-use agents interact with software through visual interfaces, browser controls, screenshots, and other representations of an application’s state. They may need to determine which button is relevant, whether a dialog is asking for confirmation, or whether a task has reached the expected result.
Text-only decision systems cannot directly inspect a screenshot unless another component first converts its contents into text or structured data. Drex Max’s ability to accept images gives developers another option: include the relevant screenshot as part of the decision request.
For example, an agent could examine a visual representation of a web form and evaluate whether the required fields appear complete. It could also classify a screen as a login page, an error message, a confirmation dialog, or a completed task, provided those outcomes are defined in the request.
This makes Drex Max potentially useful as a decision component inside a larger computer-use system. A separate agent or automation framework would still need to capture screenshots, execute mouse and keyboard actions, manage application sessions, and check the results.
The distinction is important. Visual input does not automatically give a model the ability to control a computer. Instead, it provides information that another component can use when choosing the next action.
How Drex Max Processes Images and Questions
According to the Drex Max documentation, developers select the model using the identifier drex-max in the API request. The request follows the same general structure used by other Drex models, with additional support for image inputs.
The documentation specifies support for up to eight images per request. Supported image formats are PNG, JPEG, and WebP. Images can be supplied through the documented upload process or as inline image data, subject to the relevant size and request limits.
Each question evaluates the same supplied state and images. The model then returns an answer for each question in the request. This allows a developer to assess multiple related aspects of a screen without necessarily making a separate model call for every question.
For example, an agent analysing an application screenshot might ask whether a confirmation dialog is visible, whether a particular control is enabled, and whether the interface appears to show a successful completion state. The developer defines the questions and accepted outcomes, and the model scores them.
There are limits to this approach. Nace AI notes that images are resized before inference, meaning very fine details may not be preserved. Long inputs can also encounter token or processing limits. Developers working with dense interfaces should test image quality, cropping, and the reliability of the model’s decisions under realistic conditions.
Potential Applications Across AI Automation
The ability to evaluate text and images creates several potential applications for Drex Max. Its suitability for each task will depend on the quality of the visual input, the questions developers define, and how the surrounding software handles uncertainty.
Browser and desktop automation: An agent could use screenshots to classify the current interface state and select from predefined next steps. A separate automation tool would carry out the chosen interaction.
Customer support: Systems could combine a customer’s written description with a screenshot of an error or damaged product. The decision model could score predefined categories and help route the case to the appropriate workflow.
Document and form review: A model could assess visible form states, identify whether a required visual condition appears to be met, or help classify submitted evidence. High-stakes document checks would still require suitable verification.
Quality assurance: Testing systems could compare screenshots against defined conditions, such as whether a dialog appeared or a required element was visible. This could support automated testing, although exact pixel-level comparisons may require specialised tools.
Workflow routing: Business applications could use visual context alongside written information to select a predefined process, request human review, or flag an uncertain case.
These are possible uses rather than a claim that Drex Max includes a built-in browser, desktop controller, or ready-made agent for each scenario.
Drex Max Pricing, Access, and Technical Limits
Nace AI’s documentation lists Drex Max at $0.06 per million input tokens, with output tokens not billed. The documentation states that authentication, rate limits, and billing follow the same arrangements as other Drex models.
The model supports up to 16 questions in one request and up to 255 options for a choice question or levels for a scoring question. The documented image limit is eight images per request. Input size, image dimensions, request-body size, and processing time can affect whether a request is accepted and completed successfully.
These details are useful for developers estimating costs and designing workflows, but they do not guarantee a particular end-to-end response time or operating cost. An entire computer-use agent may also incur costs from screenshot capture, other AI models, browser infrastructure, and action-execution tools.
Before deploying Drex Max in production, teams should test the model on representative screenshots and real application states. They should also set confidence thresholds, validate high-impact decisions, and ensure that actions such as submitting payments, deleting records, or changing account settings require suitable safeguards.
What Drex Max Means for the Future of AI Agents
The launch of Drex Max reflects a broader direction in AI development: not every task requires a model to generate a long response. Some applications need a reliable selection from a limited set of actions, with a clear indication of how strongly the model favours each option.
For computer-use agents, that distinction can be useful. A system may already have tools for interacting with a browser or desktop, but still needs to determine which action is appropriate for the current screen. A multimodal decision model can provide a structured way to evaluate that choice.
However, dependable computer automation involves more than selecting an answer. Agents must handle changing interfaces, unexpected errors, ambiguous visual evidence, and actions with real consequences. Developers will need to combine model outputs with application-specific rules, independent checks, and human oversight where appropriate.
Drex Max provides a new option for teams exploring that architecture. Its practical value will depend on how accurately it interprets their images and how consistently its decisions improve the complete workflow.
Final Thoughts
Nace AI’s Drex Max adds image understanding to a decision-focused model designed to return probabilities for predefined answers. That combination makes it relevant to developers exploring computer-use agents, visual classification, and multimodal business automation.
The main distinction is that Drex Max is a decision component, not a complete autonomous computer operator. It can evaluate supplied visual and textual information, but surrounding software must still manage the interface, execute actions, and verify outcomes.
For developers, the next step is to test the model against real screenshots and measurable task requirements. Strong results on representative workloads—not the multimodal label alone—will determine whether it is a useful addition to an AI agent stack.
Sources
Drex Max official documentation — model capabilities, image input, limits, API selection, and pricing.
Introducing Drex: A small model that decides instead of writing — Nace AI’s explanation of the underlying decision-model approach.
Drex official product page — product overview and published decision-model benchmarks.
Nace Document Intelligence announcement — context on Nace AI’s broader perception and decision tooling for AI agents.
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
