🚨 AI News | TestingCatalog (@testingcatalog) on X

What it claims

Drex for Finance is a decision model from Nace AI. Nace describes it as a multimodal decision model post-trained for sub-second financial decisions. Nace's announcement lists its target workloads as company filings, customer support, credit risk, live fraud prevention and quant signals. The list is cut off on the page. Nace says these workloads need "regulatory-grade decisions with zero hallucinations." That is Nace's framing, and the page gives no measurement to support it.

How it works

The input is a document or a transaction log plus a fixed set of possible answers. The output is a probability score for each answer, not free text. Nace says it trained the model directly on finance tasks, and names two examples: reading company filings and screening customers against sanctions lists. The page gives no model size, architecture, training data volume or latency figure beyond "sub-second."

Evidence

The evaluation uses 3,361 decisions drawn from 48 finance benchmarks that Nace says were excluded from training. On that set, the results are:

• Drex for Finance: 85.7% correct
• Drex Max v1.1, Nace's previous model: 77.7%
• GPT-6 Luna: 70.9%

The page does not name the 48 benchmarks or break results down by task. It also gives no calibration figures, although the model outputs probabilities.

How to use it

The model is available on the Nace Console. The page names no pricing, API shape, model identifier, weights or licence.

Limits

The source is a TestingCatalog post on X that relays Nace's announcement, and the text is truncated at "Nace has provided our…". The benchmark is Nace's own held-out set, and the page does not identify its contents. The fixed-answer format means the model ranks options supplied by the caller. Accuracy on tasks such as sanctions screening or fraud checks, taken one at a time, is not reported.