What real AI agent tasks cost
Nobody publishes this. Vendors price agent work in credits, in seats, or in monthly plans, and what one finished piece of work costs is left for you to find out afterwards. Below is every task Else has published a run for — 12 of them — with the amount charged, the engine tier that produced it, and a link to the run itself so the figure is checkable rather than asserted.
The short version
- 12 runs across 6 kinds of work: data, research, slides, travel, website, writing.
- $2.361872 in total — all 12 of them, added up.
- Median run: $0.148155 on the Smart tier, finishing in about 82.5 seconds.
- Cheapest $0.028452, dearest $0.452333 — a spread of about 16× between the smallest and largest job here.
- 2 of the 12 produced no file at all and are included anyway. Dropping them would make the set read as though every task ends in an artifact.
Download it
ai-task-costs.csv ai-task-costs.json
Published under CC BY 4.0 — use it, chart it, quote it,
including commercially, as long as you say where it came from. Suggested citation:
Else, “What real AI agent tasks cost”, 2026-08-26, https://elsedoes.com/ai-task-costs
.
Last regenerated 2026-08-26.
Every run, and what it cost
| Task | Kind | File | Tier | Time | Charged | Evidence |
|---|---|---|---|---|---|---|
| A 12-month bakery cash-flow model where the formulas are still live | data | xlsx | Smart | ~90s | $0.375000 | watch the run |
| Client proposal for an 80-page catalogue redesign, price left blank | writing | docx | Smart | ~60s | $0.131122 | watch the run |
| A one-page site for a dog walker who runs everything over text | website | html | Smart | ~60s | $0.147935 | watch the run |
| Turning three messy lines of expenses into a spreadsheet with 40 rows | data | xlsx | Smart | ~90s | $0.260554 | watch the run |
| Six weeks of fractions, no file attached, and that was the right call | writing | no file | Smart | 15s | $0.028452 | watch the run |
| Four days in Kyoto, planned around dinner, printed on one page | travel | Smart | ~90s | $0.452333 | watch the run | |
| Day-one deck for a small architecture practice, five slides exactly | slides | pptx | Smart | ~60s | $0.130120 | watch the run |
| Six quote templates for a two-van plumbing outfit, and the clause that gives it away | writing | md | Smart | ~60s | $0.148375 | watch the run |
| Six-slide investor update for a 12-seat coffee roastery | slides | pptx | Smart | ~90s | $0.203537 | watch the run |
| A scam-text sheet for a 74-year-old — right tone, wrong page count | writing | Smart | ~75s | $0.138613 | watch the run | |
| The spreadsheet a bookkeeper actually wants — on the second try | data | xlsx | Smart | ~90s | $0.209215 | watch the run |
| Three ways out of a shared Gmail support inbox, priced for six people | research | no file | Smart | ~90s | $0.136616 | watch the run |
What this dataset cannot tell you
- 12 runs is a small number. It is every run we have published, not a sample of anything. Treat it as a set of worked examples with real prices attached, not as a distribution you can generalise from.
- Every run here is on one tier (Smart). So nothing in it supports a claim about how cost varies with the engine you pick, which is the first question most people would want to ask of it.
- Each receipt carries a single line. The dataset therefore shows what a run cost, not what the run spent that money on.
- There are no dates. The published entries do not record when each run happened, so there is no time series here and no way to see prices move.
- We chose these tasks. They were written to show the product, which is a selection effect, and the honest way to read the spread is with that in mind.
- These are retail prices on one product. They are what a customer was charged. They are not a benchmark of anything else, and they cannot be compared with another vendor's per-task cost, because no other vendor publishes one.
How it was made
Each row is generated directly from the published run entry it describes — the same source
that renders the example pages — so the figure in the table, the figure on the example page and
the figure on the run's own receipt cannot disagree. A row missing its tier, its currency or its
run link fails the build rather than being published without them. The generator is
apps/marketing/scripts/build-receipt-dataset.mjs.
Hand over one job and read your own receipt
Price depends on the tier — Auto, Fast, Smart, or Max.