Route the workload. Read the tradeoff.

The sample assigns ten thousand daily requests, twelve hundred input tokens and five hundred output tokens each. Sol prices are five dollars per million input and fifteen output, Terra two and six, Luna one and three. Their per-request costs are zero point zero one three five, zero point zero zero five four, and zero point zero zero two seven dollars. At twenty thousand pixels per dollar the bars measure two hundred seventy, one hundred eight and fifty four. These are user supplied scenario prices, not current provider rates or equal quality guarantees. Routing twenty percent to Sol, thirty to Terra and fifty to Luna yields weighted per-request contributions zero point zero zero two seven, zero point zero zero one six two, and zero point zero zero one three five. Their sum is zero point zero zero five six seven. Daily spend is fifty six point seven dollars versus one hundred thirty five all Sol, saving seventy eight point three or fifty eight percent. At two pixels per dollar the blended, baseline and savings bars measure one hundred thirteen point four, two hundred seventy and one hundred fifty six point six. The month is fixed at thirty days, giving seventeen hundred one dollars. Retries, caching and outages are absent. Latency assumptions six hundred fifty, four hundred and two hundred twenty milliseconds are weighted by the same shares. Contributions are one hundred thirty, one hundred twenty and one hundred ten milliseconds, totaling three hundred sixty. At one pixel per millisecond those three bars show the contributions directly. This average is not a measured service-level promise or percentile. The application requires exactly Math.js version thirteen point two point zero and shares totaling one hundred percent. Its local library file is absent in this snapshot, so calculation and export fail closed. Native restore and reset preserve that failure rather than inventing a successful budget result.