Olypsis / Working tools
What does a useful
answer cost?
Count the calls, the failures and the work left for people. See what a correctly completed task costs.
Free to use. No sign-up.Calculations run in your browser.
Enter one average task
Count each task once, from request to final outcome. Add every model call, retry and fallback to that task’s token totals. Average across all tasks, including failed ones.
Example rates, not provider quotes. For one model and price tier, use its billed rates. For mixed models, use the token-weighted rate for each category from your bill. If you calculate model groups separately, count shared human time only once.
Count the work
Unique tasks, including failures. A retry is part of the same task.
Final results that meet your standard, including any human fixes.
All paid tool calls for that task, including retries and fallbacks.
Optional. Include time spent on tasks that still fail. Keep checking time separate from the extra time needed to fix a result.
Example estimate / USD
- Per submitted task
- $0.1260
- Per correctly completed task
- $0.1400
Where the cost goes
Excludes human time, hosting, build costs, taxes and unentered fees. This estimates cost, not output quality. The value of staff time is not automatically cash saved.
Inputs, formulas and scenario in one text file.
One change to test
What if you sent less uncached input?
Hold rates, all other tokens, tool charges, human time, task count and completion rate fixed. Real changes may also affect caching, retries, review effort and quality.
No reduction in cache reads, cache writes or output is assumed.
$1,110.00 / month in this scenario. Not measured savings.
Use the result
Your next check
Inspect the context sent at each step.
Uncached input is your largest entered cost. Look for repeated context and large tool responses, then check whether stable input is actually being read from cache.
Trace growing costsHow the estimate works
Each token category costs: tokens per task × USD per million tokens ÷ 1,000,000. Add those categories and tool charges. Use the average across all submitted tasks, including failures. Keep uncached input, cache reads and cache writes separate.
Each task includes all its attempts, restarts and fallback calls. Count it once. Correctly completed tasks = monthly tasks × final completion rate. A task repaired by a person counts once after it meets the standard. Unfinished tasks do not count as completed, but their costs stay in the total.
Optional review cost per task = average review minutes ÷ 60 × hourly labor cost. Correction cost per task = correction percentage ÷ 100 × extra minutes per correction ÷ 60 × hourly labor cost. Do not enter the same time in both fields. Include effort on failed corrections too.
Add the entered costs, multiply by monthly tasks, then divide by correctly completed tasks. If no tasks meet the standard, a cost per completed task cannot be calculated. The input-reduction scenario keeps human effort and completion fixed; a real change needs a fresh comparison.
The starting values are examples, not client results or model prices. Hosting, build costs, taxes and unentered fees are excluded. Staff time freed for other work does not automatically reduce payroll. See the worked 70/30 example and how to test cost and quality together.
From estimate to evidence
Find what is driving
your agent’s bill.
We trace the expensive steps, test the fixes, and check the result on the work your agent needs to do.