Enter last month's spend figure. This prints an itemized estimate of the portion attributable to resent, unread, and retried context — three common places inference budgets leak, and the three this model covers. Every figure is modeled from your declared number, not measured from your traffic.
Inference and API line items only. Exclude seats, training, and hosting. This page does not see your invoice; it uses the number you enter.
Sets the default waste bands below. Change them if you know your own numbers.
Marked "as entered" on the receipt. It does not attest a vendor relationship.
Estimated recoverable, monthly
—
—
Resent and unread shares apply to the input-token portion of spend, then are multiplied by a capture factor — the share of identified redundancy that can actually be pruned without changing the answer. Most resent context is load-bearing, so capture is deliberately conservative: the floor below sets the low end and the high end runs at 35%. Retries apply to total spend. The combined estimate is then capped at 45% of spend, because no honest model claims more than that from a bill it has never seen.
Nothing leaves your browser. This page loads no fonts, scripts, trackers, or images from anywhere. It makes no network requests at all. Open your developer tools and watch the network tab stay empty. Your numbers are never transmitted, stored, or logged.
Participation in an open public process is not endorsement. No agency, standards body, consortium or company listed here endorses Atom Works™, its products or its claims.