Benchmarks

prxy.monster benchmark claims should be reproducible. The public benchmark page is:

Human page: https://prxy.monster/benchmarks/
Agent-readable JSON: https://prxy.monster/benchmarks.json

Local benchmark results use synthetic fixtures, stub embeddings, and in-memory storage. They prove module mechanics. They are not a guarantee of production savings for every project.

Run the local suite

pnpm --filter @prxy/benchmarks bench

Latest local run captured on May 4, 2026:

Module	Fixture	Result
`mcp-optimizer`	Five synthetic 120-tool MCP scenarios	33.4% average tool-token reduction; 13.8% to 53.0% range
`exact-cache`	200 cache-hit iterations, static 1.8s simulated provider call	0.004ms average local cache-hit lookup
`semantic-cache`	1,000 repeated-seed synthetic queries	98.5% hit rate in that fixture
`ipc`	50%, 75%, 90% context-utilization fixtures	0% at 50%; 98.4% at 75%; 98.7% at 90%
`prompt-optimizer`	100 prefix-stability iterations	100% stable prefix on, 1% off

Run against your endpoint

prxy bench --remote -n 100

Remote benchmarking is the number that matters for your own workload. It exercises your prompts, tool catalog, cache policy, model mix, network path, and provider behavior.

What not to claim

Do not present local fixture results as customer averages. Do not claim a universal cache-hit rate. Do not claim a universal MCP reduction percentage. Quote the fixture, the command, and the date.