How the score works
The score starts at 100 and loses points for the four things that quietly inflate an AI API bill: using a bigger model than the task needs (a flagship model for simple classification can be 10–50× the cost of a small one with no accuracy gain), skipping prompt caching when the same system prompt or context gets re-sent every call, skipping the batch API for jobs that don't need a synchronous reply (typically ~50% cheaper), and retry/error waste — tokens billed for calls that failed and had to run again.
Estimated monthly waste is the score's inverse applied to your spend — a rough "what a lean setup would save," not an exact number. Use it to find the single biggest lever, then verify against your provider's actual usage dashboard.
Fix the biggest lever first: right-size the model with the blended price comparison, estimate caching savings with the prompt caching calculator, or check batch discount with the batch API savings calculator.
FAQ
How is the AI API Bill Health Score calculated?
It scores four waste sources from 100 (lean): whether your model tier matches task complexity, whether prompt caching is on for repeated context, whether batch API is used for non-urgent jobs, and estimated retry/error waste. Each unmatched or missing lever subtracts points.
What's a good score?
80+ is lean — you're using the right-sized model with caching and batching where it applies. Under 40 usually means real, fixable monthly waste: an oversized model for simple tasks, no caching on repeated prompts, or no batch discount on jobs that could tolerate a delay.
Is this exact?
No — it's a directional estimate from typical savings ranges (caching ~50-90% off cached tokens, batch ~50% off, right-sizing model tier 5-50x). Use it to find the biggest lever, then verify with your actual provider invoice.