How to Audit Your OpenAI or Anthropic Bill (and Cut It)

The fastest way to lower an AI bill is not to switch models on a hunch, it is to look at what you actually spent and where it went. Most teams skip this because it feels like a chore: export the logs, match each model name to its current rate, do the token math, then imagine the alternatives. The bill auditor does all of that in a few seconds, and it runs entirely in your browser, so a usage export with sensitive prompt volumes never touches a server.

The workflow, start to finish

The workflow is simple. Export your usage from the provider (OpenAI, Anthropic, or OpenRouter), then drop the CSV or JSON into the auditor or paste a few rows. It parses your real token counts per model, prices what you spent, and re-prices the exact same usage across every major model so you can see where you are overpaying. It then highlights a short list of like-for-like alternatives: models at or above your current capability tier that cost less, so the savings are real and not a quiet downgrade.

Read the biggest line first, and only that one

An audit produces a ranked list, and the temptation is to work down it. Do not. Spend on AI workloads is almost always concentrated: one model, one job, one runaway process carries most of the bill, and the rest is rounding. Find the row that dominates and understand it completely before touching anything else. Optimising a line worth four percent of the total is how a week disappears without the invoice moving.

When you look at that row, check the split before the size. Output tokens cost a median of 5 times what input tokens cost across the 40 models we track, so a job with modest traffic and long generated answers can outrank one with far more calls. The fix is different in each case: shorter outputs and tighter formats on one, less retrieved context and more caching on the other.

Cheaper is not the only lever, and often not the first

Switching models is the change everyone reaches for and the one that carries the most risk, because it is the only one that can alter your output. Three cheaper changes come first, and none of them touches quality.

Send less. Most prompts carry context nobody has revisited since the first version: retrieved passages set to a tutorial default, examples that stopped earning their place, a schema repeated in two forms. Cache what repeats. Where a cache-read rate is published the median discount is 90 percent off standard input, which is available on 32 of the 40 models we price and on none of the other 8. Batch what can wait. Work with no user waiting on it commonly prices at a discount on asynchronous endpoints, and evaluations, backfills and re-indexing jobs almost never need to be synchronous.

Only after those three is a model change the cheapest remaining option, and by then you will know exactly how much you need it to save.

Two caveats that keep it honest

Two honest caveats keep the numbers trustworthy. First, it compares cost, not quality. We publish no benchmark and no intelligence ranking. Where the auditor needs an ordering so it does not suggest a downgrade, it uses each provider's own output price as a stated proxy, on the reasoning that within a lineup labs price their stronger models higher. That is a proxy and it is labelled as one. A cheaper model may not do your job, and only your evaluation can tell you.

Second, the deeper cuts assume you can restructure the work to reach them. Caching pays only where the repeated part of your prompt is genuinely fixed, and batch pricing pays only where nobody is waiting. Every figure the auditor shows is a published rate or arithmetic on one, and every one is checkable against the provider's own pricing page. If ours disagrees with theirs, they are right and we want to hear about it.