DIY Jev LLM router cuts costs 9x vs Opus-only with 89% vs 93% success in 0.7s

airesearch12 · x · 2026-09-18

Developer @airesearch12 released an Auto-LLM-Router playground built on the Jev policy, with video benchmarks: same tasks run 9x cheaper than using Opus for everything, at nearly the same success rate (89% vs 93%), with routing decisions taking 0.7s — no noticeable delay.

The key insight: most routers ignore warm-cache value. A provider that already read your conversation bills the next turn at roughly 1/10 input price; switching models rewrites the whole prefix at full price. So the per-turn question isn't "which model is best" but "is a better model worth leaving a warm cache for" — and the answer changes as the cache clock runs out.

The playground visualizes the full decision: each candidate model is priced for the actual request (capability, price, measured success rate, expected total cost including retry risk and wait time), and the cheapest expected total wins. Free tiers are hidden by default since they have no cache price and break the tradeoff.

Original post →

DIY Jev LLM router cuts costs 9x vs Opus-only with 89% vs 93% success in 0.7s

More from coding & agent