Developer Dor Amir has released NadirClaw, an open-source LLM router that judges prompt complexity and automatically routes it to cheaper models, cutting API costs by 40-70% in coding and similar workflows.
Open Source · LLM Router
NadirClaw: a 10ms router that cuts LLM costs 40–70%
A self-hostable, OpenAI-compatible proxy classifies prompt complexity in ~10 ms, sending simple tasks to cheap or local models and reasoning-heavy work to premium ones — with a verifier-gated cascade that escalates when quality falls short.
~10ms
Classification overhead per prompt
40–70%
Reported cost reduction
98.3%
Quality retained vs always-Opus
1.7%
Catastrophic downgrade rate
Why routing pays off
In analyzed usage logs, most coding sessions are simple tasks that don't need a premium model — yet they're routed there by default.
55–70%
Simple tasks
30–45%
Complex reasoning
Verifier-gated cascade
Prompt inclassified ~10ms
→
Cheap / localGemini Flash · Ollama
→
Verify outputbelow threshold?
→
Escalate premiumClaude · GPT
Where it wins
Drop-in via pip install nadirclaw , localhost:8856
Works with Cursor, Claude Code, Aider, Continue, Codex
Premium quota lasted through the weekend after 2 weeks
RouterBench AUROC 0.961 · ECE 0.016 (n=11,420)
The caveats
Open version may trail the paid hosted "Nadir Pro" on accuracy
Setup requires entering each API key manually
Best value needs hardware to run local small models
Noncommercial PolyForm 1.0.0 license
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…