The Pulse
Nace Releases Drex Max for Fast Computer-Use Decisions
Nace says Drex Max can make multimodal decisions for computer-use agents in a single model pass. Its launch benchmarks show strengths on decision tests, alongside slower median API latency than its headline framing may suggest.
AI.info Team · Written with AI · fact-checked ·
Nace presents Drex Max as a way to speed up the small choices that keep computer-use agents moving, but its own benchmark table puts the model’s median API latency at 183 milliseconds per decision. That is faster than the listed GPT-6 Luna Decisions at 326 milliseconds, but slower than Drex Max’s 34-millisecond figure in a launch post on X and slower than its predecessor Drex v2.0 at 151 milliseconds. The two figures describe different measurements: the blog labels 183 milliseconds the median API latency seen by a caller, including network time, while the X post’s claim is not explained in the blog.
Announced on October 19, 2026, Drex Max is Nace’s multimodal decision model for tasks such as choosing which button to click or which search result to open. Instead of generating a written response, it scores options supplied by the developer. Nace says that can make it practical to use the model at each step in a longer task rather than sending every screenshot to a general-purpose model.
Nace bets on decisions, not generated text
“A computer-use agent spends most of its time on small decisions,” Nace writes in the announcement. The company says Drex Max takes a screenshot, screencast, scanned document, chart or photo alongside typed questions, then returns a probability for each option in one forward pass. That format is designed for questions with a defined set of answers, rather than open-ended instructions.
The Drex documentation describes a narrower, concrete API interface: developers can send text and images, with up to eight images per request and as many as 16 questions when images are included. It lists PNG, JPEG and WebP as supported image formats. The documentation does not list video as an accepted image input, so the launch post’s references to screencasts and video are not matched by the documented image-upload formats.
Strong decision scores, mixed wider results
Nace reports a score of 67.6 on its public Decision Index, ahead of Drex v2.0 at 61.3, Jev 1.14 at 60.7 and GPT-6 Luna Decisions at 47.8. On other tests in the same table, Drex Max does not lead: it scores 92.2 on JevBench, compared with 99.1 for GPT-6 Sol and 100.0 for Claude Opus 5.5. The numbers are Nace’s published results; the announcement does not describe an independent evaluation.
The results also vary across measures of agent performance. Drex Max scores 0.780 on OSWorld-Verified, above GPT-6 Sol’s 0.756 but below Claude Opus 5.5’s 0.790. The benchmark table therefore supports a narrower claim than a general-purpose model ranking: Drex Max posts the highest Decision Index score among the listed decision models, while other models score higher on some separate tasks.
Two different prices appear in the launch
Nace lists Drex Max at $0.06 per million input tokens through its API and gives the model a 262,144-token context window. The benchmark table separately estimates a cost of $0.06 per 1,000 decisions. Those figures use different units, and the per-decision estimate should not be read as a fixed charge independent of the input sent.
The company’s documentation says output tokens are not billed for Drex Max. It also says the model is selected with the pinned identifier “drex-max”; requests that omit a model use Drex v2.0 instead. That distinction matters for developers testing the new release, since a default request does not automatically route to Drex Max.
The release’s test is adoption beyond Nace’s table
Nace frames the model as a low-cost decision layer for agents that otherwise call larger models repeatedly. Its published comparisons offer evidence for that use case on selected decision benchmarks, but also show a 183-millisecond median API latency and lower scores on some broader evaluations. Developers will need to weigh those trade-offs against their own task mix and input costs.
Drex Max is generally available through the Nace API, according to the announcement. The company’s posted price is $0.06 per million input tokens; its documentation says the model accepts up to eight images in one request.