Meet Drex Max: Nace AI’s Multimodal Decision Model for Computer-Use Agents with a Score 67.59

Nace AI has released Drex Max, a decision model that reads text and screenshots and returns calibrated probabilities over options a developer supplies. It writes no text. Questions come in three forms: yes or no, pick a label, or place on a scale. Each request can carry up to eight images and 16 questions, answered in one call.

Nace reports 67.59 on the public Decision Index 0.3.2, ahead of TypeSafe's Jev 1.14 at 60.7, along with 0.780 on OSWorld-Verified and 98.2 on the BFCL tool-calling test. It lists a 174-millisecond median per screenshot decision, against 253 milliseconds for Jev 1.14 on the same questions.

Pricing is $0.06 per million input tokens with no charge for output. For agents that make thousands of small choices, that kind of pricing changes what is affordable to run.

We have been watching the decision-model category grow fast this year, and Drex Max is a sensible next step. Agents that operate a computer mostly face narrow questions: is this the right button, did the page load, should I continue. Asking a large chat model to reason in prose about each one is slow and expensive. A model that only answers from a fixed menu and attaches a probability is easier to audit and harder to talk into something off-script, because it literally cannot reply outside the options it was given.

The benchmark caveats are significant, though. Every score is self-reported. Nace says it trains on the official training splits of the index and scores on held-out splits, which is reasonable but still means the model was tuned for the very test it tops. Comparisons with other vendors' numbers are not controlled. And the gaps show the limits of the approach: Drex Max trails Jev 1.14 on general knowledge tests like MMLU-Pro, 72.5 versus 82.1, and scores 19.0 on Humanity's Last Exam. It is a fast judge, not a deep thinker.

Nace also lists three different latency figures from three different setups, from 34 milliseconds to 183 milliseconds. None is wrong on its face, but buyers should test on their own screens rather than pick the most flattering number.

The harm worth naming is what happens when per-click decisions get very cheap. Computer-use agents that cost almost nothing per action will be pointed at everything: form filling, account creation, ticket buying, scraping sites that have tried to keep agents out. A calibrated probability is only a safeguard if someone sets a threshold and stops the agent when confidence drops. At scale, the cheapest path is to let it keep clicking.

Our read: this is useful infrastructure that makes agents faster and more predictable. I'm watching for independent evaluations and for builders who publish how they handle the low-confidence cases.

Decision models, popularized by TypeSafe's Jev, return probabilities over developer-defined answers instead of generating text. Nace previously open-sourced smaller Drex models, including Drex 2.0 under an open RAIL-M license. Drex Max is proprietary, hosted only on Nace Console under a pinned model ID, with no announced open weights and no disclosed size.