Report
What an AI marketing agent task actually costs: 337 production tasks, measured
TL;DR
An AI marketing agent task is cheap at the median and expensive in the tail: across 337 production tasks the median used 349 credits (about $0.35) while the 90th percentile used 2,927. The spread comes from four things — the model, the data read, the tools called and the number of steps.
How this was measured
- Source: every completed task with a billed usage row in the 90 days to 2 October 2026 — 337 tasks across real workspaces and our own. Bring-your-own-model rows are left out, because that workspace paid its model provider directly.
- Normalised to today's prices. Each model call is re-priced from its own token counts (fresh input, cached input, output) at the 2 October price of the model it ran on, so a July task costs what it would cost today. Tool calls are converted from the credit basis they were billed at.
- What a credit is: a fixed slice of real model and tool cost. On the Starter plan a credit sells for $0.001, so 1,000 credits is $1 — the dollar figures below use that rate.
- Reproducible: the script is in the repository, and the numbers are in a downloadable CSV.
The distribution
The mean is three times the median. That gap is the single most important fact about agent pricing: most tasks are cheap, and a minority — long research, dashboards, video — carry the bill. Any per-task estimate that quotes one number is quoting the mean, the median or a guess, and they differ by 3×.
| Percentile | Credits | ≈ $ at $0.001 / credit |
|---|---|---|
| 25th | 169 | $0.17 |
| Median | 349 | $0.35 |
| 75th | 809 | $0.81 |
| 90th | 2,927 | $2.93 |
| Mean | 1,091 | $1.09 |
By kind of task
| Task type | Tasks | Median | 75th pct | 90th pct |
|---|---|---|---|---|
| Proposes or makes a change | 248 | 409 | 883 | 3,195 |
| Autonomous run (the agent picks its own next tool) | 85 | 408 | 750 | 1,971 |
| Single-platform report | 31 | 267 | 1,362 | 2,631 |
| Multi-platform analysis | 30 | 266 | 596 | 5,489 |
| General question (no platform data) | 28 | 263 | 1,234 | 1,785 |
| SEO & web research | 19 | 1,496 | 3,374 | 8,366 |
| Dashboard build or refresh | 12 | 786 | 8,061 | 9,968 |
| Video ad creative | 4 | 8,445 | 21,690 | 21,690 |
A task can be in more than one row — a multi-platform report that proposed a change is in both. Video has four runs; read its row as an order of magnitude.
What actually decides the cost
- The model, by up to 422×. The same median task costs 21 credits on GLM-4.7 Flash, 29 on GPT-6 Luna, 559 on Claude Sonnet 5, 1,397 on Claude Opus 5 and 8,872 on GPT-5.5 Pro. Nothing else in this list comes close. Every model, priced.
- How much data the task reads. The median task read 78,098 input tokens and wrote 2,450 — 32 tokens in for every one out. An agent's bill is mostly what it reads: account rows, search terms, reports, its own earlier steps.
- Paid tools. Tools were 15% of measured spend. A web search is 67 credits, a hosted-browser visit 334, an image 267; SEO research pays for every SERP and keyword pull, which is why its median is five to six times a report's.
- Caching. 43% of input tokens were served from the prompt cache at the cache-read price — the agent's instructions and tool catalogue are the same on every call, so a model with a cheap cache read pays far less for them.
How to read a vendor's "per task" number
- Ask which statistic it is. A median of 349 and a mean of 1,091 describe the same 337 tasks.
- Ask which model. A price per task without the model is not a price — the same work spans 21 to 8,872 credits here.
- Ask what a task is. A one-question report and a researched, multi-platform change are both "a task".
- A budget figure should be cautious. Our plan pages budget 1,350–1,900 credits a task when they estimate how many tasks a plan covers; 82% of measured tasks used 1,350 or fewer and 87% used 1,900 or fewer, so nobody runs out sooner than the page says.
Spending less without doing less
- Put the specialists on a cheap model and keep a stronger one for the planner, where judgment pays. The default, GPT-6 Luna, is among the cheapest models measured.
- Narrow the inputs. Select the accounts and date range the question needs; every extra row is input tokens.
- Watch the tool-heavy tasks. Research and video are where the 90th percentile lives — schedule them deliberately rather than re-running them.
Measured distributions by task type and every tool's price per call are kept current on credits by task type.
Frequently asked questions
- How much does an AI marketing agent cost per task?
- Measured on 337 production tasks: a median of 349 credits (about $0.35 at $0.001 a credit), a 75th percentile of 809 and a 90th percentile of 2,927. The mean is 1,091, because a few research and video tasks are far more expensive than the rest.
- What makes an AI agent task expensive?
- Mostly the model — the same task ranged from 21 to 8,872 credits across 49 models — then how much data it reads (input tokens outnumbered output 32 to 1), then paid tools such as web search, hosted browsers and video, which were 15% of spend.
- How much does an AI-generated video ad cost to make?
- In our measurement the four video creative tasks used a median of 8,445 credits, about $8.45 at $0.001 a credit — the coding agent that builds and renders the video is most of it. Four runs is a small sample; treat it as an order of magnitude.
- Why is the average AI agent task cost misleading?
- Because the distribution is skewed: the mean task used 1,091 credits and the median 349. A handful of long research, dashboard and video tasks carry most of the spend, so the mean describes almost no real task.