Jev sorted 154 bank support messages into 77 intents for $0.011.

Jev sorted 154 bank support messages into 77 intents for $0.011. Claude Opus 5.5 did the same 154 for $2.11 and got 83.8% right. Jev got 79.9%. What came back: → one public test, Banking77, 2 messages per intent → 0.26 s a message vs 4.2 s → accuracy, latency, cost per 1,000, raw results per message The twist: on the 122 messages where Jev said it was sure, it was right 88.5%. Higher than Opus overall. Send only the 32 unsure ones to Opus: 83.1% for $0.51. New open Jev alternatives like @TheVixhal's Gero-4b ship with speed numbers. This puts accuracy next to them, one command per model. @typesafeai Real API calls, one run, built with Claude Opus 5.5. Which model should run this test next? Repost this now, because the next team picking a classifier should see accuracy, not only milliseconds. Follow for more open source JEV builds. Code in the reply.