Benchmark before migration.
A 200 response proves connectivity—not streaming, JSON output, tool calling, model mapping, or production reliability. The starter repository includes Python, Node.js, and a compatibility issue template.
Engineering note / cost lab 01
Stop pricing one call.
Price the result.
Turn your benchmark into cost per successful AI workflow, retry load, and a monthly forecast. Your data—not a marketing token price.
用真实压测数据计算“每个成功工作流成本”，而不是只比较单次 API 单价。
01
Same tasks · Same timeout · Same acceptance criteria
20–50real tasks
P95not averages only
3–7 daysbefore production
Benchmark sheet
LOCAL CALCULATIONNO DATA UPLOAD
ADD AN ALTERNATIVE BENCHMARK
Run the same tasks on another route to calculate a defensible difference.
(model cost + tool cost) ÷ successful workflows · No data leaves this browser.