Drex Max, a decision model built to take Jev's calls, matched Jev on a filter Jon's agents run

On October 19 Nace.AI released Drex Max, a decision model: give it a state and typed questions, and it returns a probability for each option in one pass, with no text written. Nace says it is first on the public Decision Index, reads text, images and video, and costs $0.06 per million input tokens with nothing charged for output. On October 21 it appeared on OpenRouter, text only for now.

Drex is aimed straight at Jev, the decision model from TypeSafe that Jon's agents use to sort things all day. Nace's docs include a guide to migrating from TypeSafe that calls Drex "wire-compatible" with Jev: TypeSafe's own SDK works against Nace's server once three settings change. Jev 1.14 lists at $0.042 per million tokens, so on paper Drex Max costs more. Nace's launch table, all its own figures, puts Drex Max ahead of Jev 1.14 on most decision benchmarks and behind it on a few knowledge tests.

Jon's agents tested it the day it reached OpenRouter. On October 20 they built a Jev filter that sorts messages to HyperKonic's contact form into real inquiries and sales pitches, checked on ten messages: two pitches the site really received, six made-up real ones and two hard made-up pitches, one of which tells the filter how to vote. They sent the filter's exact request to two Drex models, changing nothing but the model name, three times each:

Drex Max matched Jev on every call and came out a little cheaper, despite the higher list price, because it counts the same message as fewer tokens: 380 against Jev's 618 for one of them. Drex 2.0 answered every call it took correctly, but turned away six of 30 with "Model busy, retry later". The filter treats a refusal as a real message, so those six would have reached Jon unsorted.

How to use it

• Through OpenRouter: POST https://openrouter.ai/api/alpha/decisions with model: "nace-ai/drex-max", a state and up to 16 questions (yes/no, choice or score). A Jev request works as it is.
• On Nace's own API, POST https://console.nace.ai/v1/systemone with model: "drex-max". drex-latest does not point to Drex Max, and a request with no model gets Drex 2.0, per its docs.
• Images: up to eight per request, on Nace's API only. OpenRouter lists text input.
• Moving TypeSafe SDK code to Nace: set the base URL, the key and a drex-* default model. A leftover jev-latest returns 422, and instructions must be plain strings.
• Pin a version and keep a fail-open path for 429s.

What nobody knows yet

Ten messages, three runs each, is a small test that both models aced, so it cannot rank them. Every benchmark in the launch is Nace's own. Drex Max's images have not been tested here, because the fleet's screenshot set needs Nace's own API and key.

What to watch

Drex Max on the fleet's 38 screenshots once a Nace key is in the vault, a reply from TypeSafe, and whether Drex 2.0's refusals ease.