Learn · AI for leaders
What does AI cost to run?
Chat seats cost a flat monthly fee, while automated workflows pay by the page read and written. Split a workflow between a large model and a small one, and watch the monthly bill.
Free lesson, about five minutes. Updated
- Step on large model
- Step on small model
- All on large
- $4,500a month
- With split
- $4,500a month now
- One task
- 45.0¢5 large, 0 small
- Saved
- $0$0 a year
-
1 PredictWhat happens when every step uses the biggest model?
An operations team runs an AI workflow 10,000 times a month. Each task reads 20 pages of documents across 5 steps: sorting, pulling fields, checking dates, formatting, and writing one judgment call. What does the monthly model bill look like if all 5 steps use a large frontier model instead of routing the 4 routine steps to a small model?
Pick one first. A guess that turns out wrong teaches the most.
Locked in. Now test it.
-
2 TryMove the routine steps to a small model
Move Steps on small model from 0% to 80% (4 of the 5 steps), and watch the bars and the monthly bill.
- Monthly bill
- $4,500
- Saved a month
- $0
Use your own numbers
-
3 ObserveWere you right?
Commit to a prediction first.
-
4 ExplainWhy does it come out this way?
Automated workflows pay by the token, roughly three quarters of a word, every time a model reads a page or writes a reply. When an agent runs a five-step loop over a 20-page packet, it reads those 20 pages five times.
A large frontier model costs roughly fifteen times as much per page as a small model. Using the large model for the one step that calls for judgment, and a small model (or plain code) for sorting, pulling fields and formatting, cuts the bill by three quarters without hurting the result.
-
5 ApplyA new case
Your team builds an internal search tool over 50,000 company PDFs. The first prototype sends the top 30 full PDFs (600 pages in all) to a frontier model on every employee question, and the bill hits $18,000 in the first month. What do you fix first?
Rationing questions makes the tool less useful while leaving every question just as wasteful. Fix what each question reads first.
Right. Model bills follow the number of pages sent in. Indexing documents by section and sending only the 5 pages that match the question cuts the pages read per ask from 600 to 5, reducing the bill by more than 95% and giving faster, sharper answers.
Per-seat chat plans enforce their own context and rate limits behind the scenes, and still need retrieval over 50,000 PDFs to find the right pages.
The math behind the figure
- One page of text is taken as about 500 tokens (roughly 375 words), including the prompt and a short structured reply at each step.
- The large frontier model is priced at $4.50 per 1,000 pages read and answered (equivalent to $9 per million tokens blended), or 9 cents for a 20-page step.
- The small model is priced at $0.30 per 1,000 pages (equivalent to $0.60 per million tokens blended), or 0.6 cents for a 20-page step, one fifteenth the cost.
- One task costs the large-model steps plus the small-model steps. At 5 steps on 20 pages, all large is 45 cents a task; 1 large and 4 small is 11.4 cents a task.
- Prompt caching and plain code lower the bill further when the same 20 pages are reused across steps or when a step can be done by a script with no model at all.
AI for leaders
Five lessons on what AI is, where the money is, and what has to be true first.
- 1Why does the same question get different answers?5 minutes
- 2What is one workflow worth?5 minutes
- 3Why can’t AI answer questions about your own business?5 minutes
- 4Where does your data go when you ask AI?5 minutes
- 5What does AI cost to run?This lesson