Drex Max API: Pricing, Computer Use and Context Limits

Nace.AI released Drex Max on October 19, adding image-based decisions to its hosted decision-model API. It takes a screenshot or other input and scores options supplied by the developer, rather than writing an answer. For computer-use agents, that offers a different way to handle repeated decisions inside a longer task.

The launch announcement describes the model as generally available, with support for screens, documents and video-based workflows. The important distinction is what comes back: probabilities over choices, rather than an open-ended response.

What Drex Max actually does

Imagine an agent looking at a checkout page. A general-purpose model might explain the page, generate a plan and then propose an action. A decision model can instead score a supplied set of actions, such as continue, edit the address or ask for help.

That can make the interface easier to validate. The application knows the available answers before making the request and can reject or escalate a decision that does not meet its confidence threshold.

It also moves responsibility into the application. The right action has to be represented among the options. A confident answer from an incomplete choice set is still a bad decision. The developer needs a way to express uncertainty, refresh the available actions and handle situations the workflow did not anticipate.

Drex Max pricing

Nace's pricing documentation lists Drex Max at $0.06 per million input tokens. Output tokens are not billed, and there is no separate per-request or per-question charge. Image tokens count toward input usage.

At that rate, 100 million billed input tokens would cost $6 for model inference. That is an arithmetic example, not an estimate of how many browser tasks the model will complete. Screen complexity, request construction and repeated attempts all affect consumption.

The distinction matters because Nace also offers managed browser runs. Those have separate charges for Drex Max usage, a planning and reading model, and browser time. The direct decision API price should not be presented as the complete cost of running a browser agent.

How to select Drex Max in the API

The Drex Max guide specifies the pinned model ID drex-max on POST /v1/systemone. Using drex-latest, or leaving the model unspecified, does not select Max: the documentation says the default remains Drex v2.0.

Requests combine shared state with named questions. Max supports up to 16 questions per request and up to 255 options for a choice question. Images belong in the dedicated images array, with a maximum of eight per request. A URL included as ordinary text is not fetched automatically.

Those details matter more than a generic claim of multimodal support. An existing integration will need to select Max explicitly and supply image data through the supported path. A screenshot pasted into a text field is not the same thing.

The context window needs two numbers

The launch advertises a 262K-token context window. The API limits reference makes a more useful distinction: Drex Max has a 131,072-token state limit and a 262,144-token limit for the state plus the longest question.

So developers should not assume they can submit a 262K-token document as state. The Max guide also says long state may be shortened before inference, and long multi-question requests can hit shared limits or time out. An accepted request is not a promise that every part of a long document was used unchanged.

For a screenshot-driven workflow, it makes sense to test a compact representation of the current screen before filling the available context. Old page content can distract a decision model just as it can distract a generative one.

What the benchmarks do not settle

Nace reports a median API latency of 183 milliseconds and a Decision Index score of 67.6 for Max. These are vendor-reported results, not independent measurements of a particular production workflow. Its own table also shows Max behind Drex v2.0 on When2Call, despite improvements elsewhere.

The Drex product page distinguishes Max from the open-weight Drex 2.0 model. Its 35B-total, roughly 3B-active parameter description applies to Drex 2.0. It should not be reused as a verified size claim for Max.

The useful evaluation is whether Max makes the right small decisions cheaply enough, with an escalation path when it cannot. A lower inference bill does not compensate for an agent confidently choosing the wrong button. For developers, the opportunity is to separate planning from repeated choice and measure both parts independently.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI founders turn technical launches and product decisions into clear stories that developers and founders actually read and share.

Book a 30-minute call.