Nace releases Drex Max vision decision model
Nace has made Drex Max generally available, a multimodal model that scores the small step-by-step choices in computer-use agents in milliseconds, aiming to replace slow, costly frontier-model calls inside agent loops.
Nace announced the general availability of Drex Max, which it calls its most capable decision model and the first in its Drex line that accepts images and video as well as text.
Drex Max is not a generative model. The caller supplies a screenshot, screencast, scanned document, chart or photo with a set of typed questions, and one forward pass returns a probability for each option. Nace says this fits the many small choices a computer-use agent makes, such as which button to click or whether a page has loaded. A single task can involve around thirty of these, and today each one often goes to a large frontier model.
In Nace's own benchmarks, Drex Max scores 67.6 on the public Decision Index 0.4.1, the top result among the decision models it compared. On OSWorld-Verified it scores 0.780, against 0.790 for Claude Opus 5.5 and 0.756 for GPT-6 Sol. It trails both of those models by a wide margin on knowledge tests such as HLE (19.0, against 56.0 for Opus 5.5).
Nace reports a median API latency of 183 ms per decision and a cost of $0.06 per 1,000 decisions. For comparison, it lists about 3.5 s and $10.17 for Claude Opus 5.5. On its vision index, Drex Max scores 74.5 against 63.9 for GPT-6 Luna Decisions.
The model is available through the Nace API at $0.06 per million input tokens, with a 262K-token context window. All benchmark figures come from Nace and have not been independently verified.