Free tool · runs in your browser

What does your AI workflow really cost?

Most AI cost calculators stop at the model's bill. This one adds the people who check the AI's work, and the mistakes that get out unchecked. See the cost per 1,000 records, what a better model would have to save to pay for itself, and when a review is worth it.

Try:

01 · The bill

What do 1,000 records cost?

A record is one item through the workflow: an article, a ticket, a document. Every share below is of all records that enter, duplicates included.

The model

Run as a batch: 50% off, results within hours instead of seconds

Tokens per call, counting everything sent (system prompt included) and everything returned, thinking tokens included: they are billed as output. Your provider's usage report or the API response's usage field has the real numbers. Calls per record is an average: 0.6 for a step 60% of records reach, 3 for a loop, 0.04 for one call per batch of 24.

The people

Sent to a person Minutes per review Hourly cost of a reviewer

Use the full cost of an hour, not the wage. In the US, wages are 70% of what employees cost employers (BLS, June 2026), so divide a wage by 0.70. Minutes per review is the whole handling time: open, read, decide, fix, move on.

Volume

Records per month

More options: blocked records, fixed costs, doing it by hand

Blocked or suppressed as duplicates Never reaches a person or a customer. Fixed costs per month Platform subscription, hosting. Minutes to do one by hand With no AI at all, to compare.

02 · Model choice

Would a better model pay for itself?

A better model costs more per token, but if it's right more often, fewer records need a person. How many fewer does it take to cover the price? Uses the workflow you entered above.

Change the model in Switch to

Its input tokens Its output tokens

Per call. Blank means the same as now. A model that thinks more writes more output tokens.

Its share sent to a person, if measured

03 · Review threshold

When is a review worth it?

Reviews cost money, and so do mistakes that slip out to customers. This answers three questions: when one review is worth it, whether your review queue pays for itself, and what the mistakes that still get out cost you.

What does one mistake cost you?

When a wrong result reaches a customer: a refund, rework, a complaint, a lost client. A rough guess is fine; the results show how much rides on it.

Of the mistakes reviewers see, how many do they catch?

Your test results, checked against an answer key

Records that went out without review How many of those were wrong

Of the records sent to a person, how many really needed one?

This one is called review precision. workflow-eval-harness measures all three test numbers.

Prices · editable

Model prices

Model	Input	Cached input	Output	Notes

Checked, not assumed

How do you know the numbers are right?

A cost calculator people make decisions with should prove its arithmetic. This one is checked three ways, and every check is in the repository for anyone to run.

25 of 25 worked-example figures match Every number in the method's worked example, recomputed with exact decimal arithmetic.

80,000 values match a second implementation 5,000 random workflows, built to hit every edge, run through a separate exact version of every formula.

14 of 14 deliberate mistakes caught The tests were run against broken copies of the code, such as minutes read as hours, or a price change applied a day early. Every copy failed.

$13.82 reproduced to the cent A real workflow's old cost estimate per 1,000, from its token counts. It used an out-of-date price: at today's prices it is $9.21, which the workflow now uses.

How it works

The math, in plain words

Model cost

Tokens in and tokens out, each at the model's price per million, for every call a record needs. Cached input is billed at the cache price, batch work at half price, and prompts over a model's long-context threshold at its higher price.

People cost

The share of records sent to a person, times the minutes each review takes, times what an hour of a reviewer's time costs. It is usually most of the bill: a review costs dollars, and a model call costs fractions of a cent.

Would a better model pay?

Every point of records sent to a person has a price. Divide the extra model cost by it, and you get how many points fewer the better model must send to a person to break even. This assumes the new model's mistakes are no more frequent.

When is a review worth it?

A review is worth it when a record's chance of being wrong, times what a mistake costs, is more than the review costs. A $2.68 review and a $50 mistake break even at a 5.4% chance. That is the cost-saving rule Chow worked out in 1970. Using it on your AI's confidence scores only works if those scores are honest (calibrated).

Mistakes that slip out

Your test's count of mistakes is a sample, so the page shows its likely range (the same 95% range as the error-rate calculator), carried through to mistakes per 1,000 records and what they cost. The cost depends on three things: how often unreviewed records are wrong, how many go out unreviewed, and what one mistake costs.

What it leaves out

Building and maintaining the workflow, reviewer training and idle time, other per-record charges, and taxes. The method lists every assumption and when it breaks.