Open-Source LLM Router Brick Cuts Costs by Up to 22x
alexgenovese · reddit · 2026-07-06
The Regolo team open-sourced Brick (Apache 2.0), a hybrid model router. By classifying capabilities (coding, math reasoning, creative, planning, world knowledge) and evaluating complexity for each prompt, it automatically routes queries to the cheapest yet competent model in the pool. Decisions are made in a single forward pass without cascade waste. Tests show it can reduce LLM costs by 4.7x to 22x. Taking 1,000 daily calls as an example, an original monthly fee of $4,950 can be drastically reduced after routing. The project is open-source and supports OpenAPI-compatible interfaces.
More from Infra
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27
- Thieves spoofing GPS steal $100M+ in data center cargo since June — Ars Technica AI · 2026-10-27
- Lam Research: weekly token usage up ~250x in 18 months — Beth_Kindig · 2026-10-27
- AWS shows how to build enterprise agentic retrieval on Bedrock with CDK — AWS ML Blog · 2026-10-27
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27