299 real intents tested: classifier routing trails GLM-4-Flash by 3 points but is 6.5x faster

Sufficient_Flower860 · reddit · 2026-09-23

The author benchmarked two routing approaches on a production Feishu life-assistant intent router, using 299 real user intents with the same gold labels and a hard 19-skill suite:

Failure modes differ sharply: 23 of the 29 GLM-only cases were fuzzy skill boundaries (trip-planner vs life-service), while 15 of the 20 Jev-only cases were GLM JSON/parse failures that fall back to "unknown" in production.

Performance and cost: Jev p50 0.55s vs 3.6s; format parse failures 0 vs 16; the full 299×2 run cost $0.086.

Bottom line: the author still puts the classifier first on the critical path, but Jev only routes — it doesn't extract params, so production still needs a generative step behind the gate. Measured with llm-fit on the actual production intent path, not a public leaderboard.

Original post →

More from coding & agent

coding & agent channel →