AI cost worksheet and decision log

How to use these tools

Analyze your billing exports to review reported costs. Use the worksheet below to record one workload, then use the decision log to plan and review a change.

Notes stay in this tab and clear on reload. Download a copy before closing or reloading the page. Nothing entered here is sent to an analysis service.

01 / Record costs

Record costs for one workload

Record one workload for one billing period using invoices and usage logs. This worksheet records notes; it does not calculate totals. Short notes fit the one-page print layout; longer notes may take another page.

Workload and user taskName the feature or workflow. What counts as one user task before retries?

Name the feature or workflow. What counts as one user task before retries?

Billing period and person responsibleRecord the dates covered and who will review these costs.

Record the dates covered and who will review these costs.

Sources, task counts and token usageLink the invoice or usage log. Record task counts and average uncached input, cache-read, cache-write and output tokens. Mark estimates.

Link the invoice or usage log. Record task counts and average uncached input, cache-read, cache-write and output tokens. Mark estimates.

Cache usage and pricesWhich input is reused? Record the cache hit rate, the request or token count used to calculate it, read/write prices and expiry. What could explain misses?

Which input is reused? Record the cache hit rate, the request or token count used to calculate it, read/write prices and expiry. What could explain misses?

Retries and chargesRecord attempts per user task, whether retries are charged, the source and the main retry reason. Write unknown where needed.

Record attempts per user task, whether retries are charged, the source and the main retry reason. Write unknown where needed.

Other fees and costs not includedRecord additional charges, credits, taxes, currency and any costs excluded from your totals.

Record additional charges, credits, taxes, currency and any costs excluded from your totals.

Task success and answer qualityRecord accepted tasks divided by original user tasks, how quality was checked, and evidence of any change.

Record accepted tasks divided by original user tasks, how quality was checked, and evidence of any change.

Response timeRecord the median and 95th-percentile (P95) response time, the source and your maximum acceptable wait.

Record the median and 95th-percentile (P95) response time, the source and your maximum acceptable wait.

Engineering timeRecord hours spent, what the work covered, and whether twice the effort would change your decision.

Record hours spent, what the work covered, and whether twice the effort would change your decision.

Total cost and cost per successful taskRecord token charges, other fees and total cost. Divide by original tasks and accepted tasks separately, showing the task counts used.

Record token charges, other fees and total cost. Divide by original tasks and accepted tasks separately, showing the task counts used.

Expected cost versus reported costCompare your expected and reported costs. Check request volume, input, output, cache, retries and fees. What remains unexplained?

Compare your expected and reported costs. Check request volume, input, output, cache, retries and fees. What remains unexplained?

Next check and review dateWhat needs checking next? Record who will check it, when, and what result would change the decision.

What needs checking next? Record who will check it, when, and what result would change the decision.

Do not count cached input twice. If your export includes cached tokens in total input, subtract them to get fresh input. Reasoning already included in output stays in output.

02 / Plan a change

Plan a change and track the result

Plan one change at a time. Record the options, measurements needed and review date before the test. Add the measured result after the review.

DateWhen the decision was made.

When the decision was made.

Change to testDescribe one change so you can measure its effect.

Describe one change so you can measure its effect.

Options consideredInclude keeping the current setup.

Include keeping the current setup.

Measurements and sourcesRecord the billing period, invoice, usage log, quality test or experiment. State what was measured.

Record the billing period, invoice, usage log, quality test or experiment. State what was measured.

Unknowns, risks and assumptionsList missing measurements, possible risks and assumptions to check.

List missing measurements, possible risks and assumptions to check.

Person responsibleName the person who will run the test and review the result.

Name the person who will run the test and review the result.

Review dateChoose when you will check the measured result.

Choose when you will check the measured result.

Measured result after reviewRecord the result, how and when it was measured, and its source. Keep predictions separate from measured results.

Record the result, how and when it was measured, and its source. Keep predictions separate from measured results.

Keep, adjust or undo the changeExplain the next decision. If measurements are still missing, write unknown and set a new review date.

Explain the next decision. If measurements are still missing, write unknown and set a new review date.

If the review arrives without evidence, write unknown and set another review date. Do not turn a forecast into an outcome.

Related tools and blank templates

Compare prices or analyze your bill

Compare published model prices and estimate a workload, or analyze your billing exports. Download a blank Markdown template if you prefer to take notes in another app.

Compare model prices Analyze billing exports Download blank cost worksheet (.md) Download blank decision log (.md)