Nace’s Drex Max puts a decision model in the computer-use agent’s hot path

Nace has launched Drex Max, a model that reads screens and chooses from options supplied by an agent, rather than writing a chat-style answer. OrcaRouter reports that Nace is pitching it as a faster, cheaper way to handle the repeated decisions in computer-use tasks such as navigating a booking site.

Watch Desk analysis

What happened

OrcaRouter says Drex Max became generally available on 19 October. The model takes a state description, questions and up to eight images, then returns probabilities for defined choices, yes-or-no questions or ordered scores. It has no open weights: Nace offers it through its API and also lists deployment on dedicated infrastructure, in a customer’s AWS or Azure account, or in an air-gapped setup.

Nace’s figures, as reported by OrcaRouter, put the price at $0.06 per million input tokens, with no charge for output. Its table gives a 183-millisecond median latency across a benchmark suite. A separate screenshot test is reported at 174 milliseconds, so these are different measurements, not interchangeable headline numbers. Nace’s comparison with other models is its own, unaudited benchmark work, not an independent test.

Why it matters

A computer-use agent may need to decide where to click, whether a page has loaded and which result to choose many times in one task. A model that returns probabilities for actions the developer has already listed could fit that loop better than a general-purpose model asked to explain its choice in prose. The software can set a confidence threshold, send uncertain cases to a slower model or person, and avoid parsing free-form answers for actions that were never offered.

That is a specialised tool, not a new all-purpose brain. OrcaRouter’s account says Drex Max trails frontier chat models on some general-knowledge and reasoning tests. It also reports an inconsistency in the published context limit: the API guide says 131,072 tokens per question, while the launch post says 262,000. Developers sizing requests should treat that discrepancy as an open question, not pick whichever number looks friendlier.

Our read

The interesting bet is architectural: use a small, constrained decision step for the many routine choices in an agent, and save open-ended generation for the parts that need it. That could make a computer-use workflow cheaper and easier to control. The catch is that the appealing speed and cost comparisons come from Nace’s own testing, and the model is hosted rather than downloadable. Useful plumbing, with the usual need to test it on the workload that actually matters.

What to watch

• Whether independent developers reproduce Nace’s cost, latency and benchmark results.
• Whether the API documentation reconciles the two published context limits.
• How Drex Max performs in complete agent workflows, where screenshots, page waits and task completion matter too.

Discussion spark: Would you trust a model that only chooses among actions you define for routine computer tasks, or do you need a general-purpose model to explain and adapt its choices?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.