Checking a new rival decision model against Jev on a filter that is already live
I took the exact Jev request a company contact form uses to sort real inquiries from sales pitches, and sent it unchanged to Nace.AI's Drex Max and Drex 2.0 on OpenRouter, changing only the model name. Ten messages (two real pitches the site received, six made-up real inquiries, two hard made-up pitches, one of which tells the filter how to vote), three runs each. Jev 1.14 got 30 of 30 at a median 192 ms. Drex Max got 30 of 30 at 302 ms and cost a little less in total ($0.00070 against $0.00079), because it counts the same message as about 380 tokens where Jev counts about 618. Drex 2.0 answered all 24 calls it took correctly but refused 6 of 30 with "Model busy, retry later".
What I learned
On OpenRouter's decisions endpoint a Jev request runs on Drex as it is, so trying a rival is a one-word change. Compare cost per call, not list price: token counts differ by model. Keep the fail-open path for 429s; a busy model is a different failure from a wrong answer. Ten easy cases can't rank two models that both ace them; it takes a harder set.
status: measured · questions: one choice: inquiry or sales_pitch · result: Jev 1.14 30/30, Drex Max 30/30, Drex 2.0 24/24 answered + 6 refused (429) · where it lives: contact-form pitch filter