Hey PH! I'm Dor, the maker of Nadir. I built this because I was spending $800/mo on Claude Opus for an app where half the prompts were simple classifications and reformats. Switching those to Haiku manually meant maintaining routing logic that broke every time I changed a prompt. So I trained a classifier to do it automatically. The key insight was adding a verifier, a second, lightweight check that catches when the classifier routes a hard prompt to a cheap model. That's what lets us promise 98% quality: the verifier is the safety net. Where we are today: 60% average cost savings on real workloads #3 on RouterArena, 92.1% on RouterBench (vs 27% for Not Diamond on the same set) 180ms routing overhead on CPU Two lines to integrate — works with every OpenAI-compatible SDK Would love your feedback, especially on routing quality. Try it on your weirdest prompts — that's where routers usually break, and where we've spent the most time.

Nadir

Cost Aware Model Rotuer

Cost Aware Model Rotuer