Drex Max turns decision models into screen readers: 67.6 on the public Decision Index at 183 ms per decision

A computer-use agent makes many small decisions per task, such as which button to click, which field to fill, which search result to open, or whether a page has finished loading, and a single task can take thirty of them. Drex Max turns each of those into a score over supplied options rather than a written answer, and it is the first decision model that reads images and video. The prior pattern sends a screenshot to a large frontier model at each step and waits seconds for it to write an answer; the earlier Drex v2.0 and v1.5 had no visual input. On the public Decision Index 0.4.0, Drex Max scores 67.6 against 61.3 for Drex v2.0 and 58.1 for Drex v1.5, with Jev 1.14 at 60.7 in the same table; on the vision side the public Vision Index is 74.5 against 63.9 for GPT-6 Luna Decisions.

It writes no tokens, so a decision costs one pass over the input no matter how many options are scored, and cost does not grow with output length. Generative models pay for every token they write, so more options and longer explanations cost more; Drex Max fixes the cost at one forward pass. Median API latency as the caller sees it, network included, is 183 ms for Drex Max, about 2.0 s for GPT-6 Sol and about 3.5 s for Claude Opus 5.5; cost per 1k decisions is $0.06, against $3.33 and $10.17 for those two.

The report covers Nace's release information for Drex Max and its accompanying benchmark tables: decision, vision, knowledge, reasoning, agentic and efficiency metrics, plus availability through the Nace API at $0.06 per million input tokens with a 262K-token context window. The flight-booking and live-camera demonstrations are recordings supplied by the releasing company, with no independent reproduction reported.

The releasing company states that Drex Max ranks first among decision models on the public Decision Index 0.4.0, but that index is a public leaderboard and the table gives no Decision Index value for GPT-6 Sol or Claude Opus 5.5. The table also shows 85.2 on When2Call, below Drex v2.0 at 86.5 and Drex v1.5 at 89.7, and slightly below its respective comparators on KIE and Hateful Memes.