Most AI bills mix several models into one total. That total hides the single largest optimization most teams have never made: matching each task to the least expensive model that handles it correctly. You cannot make that call without first tracking spend by model.
Why the model breakdown matters
The price difference between a frontier model and a smaller one for the same token volume is often an order of magnitude. When every request routes to the top model regardless of difficulty, the routine work is being charged at premium rates. Tracking spend by model is what makes that overspend visible.
How to break spend down by model
Model attribution is usually the easiest breakdown to get, because the model name is already present on every API response. Group your usage data by model to see what share of total spend each one drives, then cross it with call volume to get average cost per call per model. That pairing tells you where the expensive models are doing routine work.
What to do with the breakdown
When a large share of spend sits on your most capable model, ask what fraction of that work genuinely needs it. Take a sample of those requests, run them through a smaller model, and score the results against your quality bar. The requests where the smaller model matches are safe to route down, often cutting the cost of that workload dramatically with no change in output.
Keep tracking as models change
Model economics shift constantly. A task that needed a frontier model last year often runs fine on a smaller one today. Tracking spend by model is not a one-time audit; it is an ongoing view that tells you when a cheaper model has become good enough to take over a workload.
The model breakdown is where the fastest large savings on an AI bill usually hide. Track it, and right-sizing becomes a routine decision instead of a rare project.