Token Waste Calculator
How many of your tokens actually contributed to output? Enter your four pillars and see estimated waste, efficiency, and a breakdown by category.
Enter your four token pillars
From ccusage --json or npx sigrank me. Token counts only.
Waste breakdown by category
Input beyond a 10:1 input-to-output ratio — likely over-specified prompts.
Cache-write tokens never read back — context cached but not reused.
Cache-read tokens far exceeding what plausibly shaped output — re-sent boilerplate.
This is a heuristic estimator, not a measurement. The categories use rough proxies (a 10:1 efficient input-to-output ceiling, cache-write vs cache-read reuse, a 20:1 cache-read-to-output threshold). Real waste depends on your actual prompts and workflow. Use the numbers as a directional signal, not an audit. The Υ Yield metric is the rigorous counterpart.
What “waste” means here
A token is “wasted” when it likely did not contribute to useful output. The three categories the calculator surfaces — excess fresh input (over-specified prompts), unreused cache writes (context cached but never read back), and repeated context (boilerplate re-sent every turn) — are the most common ways operators burn tokens without compounding signal.
This is a heuristic estimator. Token counts alone cannot see prompt content, so the category thresholds are approximations. Treat the numbers as a directional diagnostic, and pair them with the rigorous Yield Calculator for the precise efficiency metric.
Frequently asked questions
What is AI token waste?
Token waste is the share of tokens that did not plausibly contribute to useful output — over-specified prompts, context cached but never reused, and boilerplate re-sent every turn. Reducing waste raises your yield and lowers your cost without changing the model you use.
How does the token waste calculator estimate waste?
It uses three heuristic proxies: excess fresh input beyond a 10:1 input-to-output ratio, cache-write tokens that are never read back, and cache-read tokens far exceeding what plausibly shaped output. The three are summed into an estimated waste total and an efficiency percentage. These are rough proxies, not a measurement.
Is the waste breakdown accurate?
No — it is directional, not exact. The category thresholds (10:1 and 20:1) are approximations. Real waste depends on your actual prompts and workflow, which token counts alone cannot see. Use the breakdown to spot where to investigate, then use the Υ Yield metric as the rigorous counterpart.
How do I reduce wasted tokens?
The highest-leverage fixes are prompt caching (so context is read, not re-sent), trimming repeated boilerplate from prompts, and writing denser output-bearing prompts instead of long open-ended ones. The waste breakdown shows which category dominates for you.
How is this different from the yield calculator?
The yield calculator computes the rigorous Υ Yield metric (cache_read × output / input²). This tool estimates waste with heuristic category breakdowns — useful for diagnosing where tokens are being burned, but less precise than yield. Use both together: yield to track, waste to diagnose.