How much of your AI bill is avoidable?
Enter one month of token usage from your provider's dashboard. You'll see what four common changes could save, worked out from list prices. It takes about a minute.
An estimate from list prices, not a measurement. Prices are each provider's standard list prices, checked 10 November 2026. Real savings depend on your traffic, and every change except batch pricing needs a quality check before it ships. That check is what an engagement does.
The calculator runs in your browser. Nothing you type leaves it unless you choose to email it.
Calculator
Scripts are off in this browser, so the calculator can't run. You can still send your usage totals for a free written estimate.
| Change | Saves a month | Before it ships |
|---|
Changes are applied in the order shown, so savings don't double count. Caching assumes 80% of the repeated part is served from cache. Cache write fees are left out.
The email opens in your own mail app with these numbers filled in. You get a written estimate within 48 hours, at no charge. How the bill check works.
A cheaper setup that answers worse isn't cheaper
In Teardown #1, the fastest and cheapest setup on one GPU quietly cut answer accuracy from 83.6% to 79.6%. It would have looked like the winner without a quality check. Routing easy requests to a smaller model has the same risk, so it's tested on your real requests before anything ships.
The useful number is cost per successful answer, not cost per request.
- Prices
- Standard list prices from OpenAI, Anthropic and Google, checked 10 November 2026
- Batch
- 50% off list on all three providers, results within 24 hours
- Cheaper model
- The smallest model from the same provider
- Not included
- Cache write fees, long-context surcharges, volume discounts