Open-source JEV router cuts LLM costs from $0.24 to $0.0001 per query with 300ms routing
1337NET · reddit · 2026-09-21
- A Reddit user built a CLI around TypeSafe AI's JEV, a small router model that sits in front of expensive LLMs: it reads the question first and decides which model answers.
- Real runs: "capital of Portugal" routed to Haiku at $0.0001, while "multi-region postgres failover" routed to Opus at $0.2434.
- Routing costs 300ms and 200 tokens per question; it doesn't need to be right every time, just often enough to beat that overhead. On failure (timeout, bad JSON) it falls back to a default tier so the request never dies.
- Written in TypeScript with zero dependencies, MIT-licensed on GitHub (adityaarakeri/jev-router). The author asks about non-model-call routing approaches like embedding similarity or length heuristics.
More from coding & agent
- Dev combines Jev and BaoCut into a local tool that finds video clips from one sentence — dotey · 2026-09-21
- claude-ops turns Claude Code into a business OS with 57 skills, 21 agents — tom_doerr · 2026-09-21
- Inside OpenAI's agentic software factory: Codex takes over, IDEs and pull requests fade — AxSaucedo · 2026-09-21
- Developer Gives AI Agent a Phone, Turns It Into a Personal Concierge — ethanniser · 2026-09-21
- Benchmarks show coding agents edit code they shouldn't in 35-65% of cases; prompt framing is the lever — RunAI_Coder · 2026-09-21
- Using a second LLM as a watchdog to catch coding agents faking success — Ascend-910 · 2026-09-21