What an AI employee costs per month (our real numbers)
In October 2026, nine of the AI employees we run cost more than a dollar in model fees, and six of those landed between $10 and $60 for the month. The busiest one cost $299. Model fees are what the AI companies (Anthropic, OpenAI) charge for the thinking each one does.
These come from our own billing records, not a spreadsheet model. 17 AI employees and helper agents ran in October. All 17 together cost $485.
The October numbers
A run is one piece of work: a reply in chat, or one scheduled job like a morning brief.
| AI employee | Runs | Model fees |
|---|---|---|
| The busiest one (a client's) | 3,556 | $298.88 |
| My own chief of staff | 921 | $56.01 |
| My bookkeeper | 121 | $12.55 |
| My inbox agent | 61 | $1.73 |
Of the nine that cost more than a dollar, the one in the middle cost $19.65.
Chat costs more than routines
| Kind of work | Runs | Model fees | Median run |
|---|---|---|---|
| Scheduled jobs (briefs, checks, reports) | 2,855 | $198 | 4.6 cents |
| Replies in Slack | 1,079 | $164 | 12 cents |
| Replies in the web app | 585 | $94 | 7 cents |
Median means half the runs cost less and half cost more. The median Slack reply cost about two and a half times the median scheduled job. Part of the reason: with each chat reply, the conversation so far goes to the model again, so a long thread costs more per reply.
The busiest one shows it well. Its 2,446 scheduled runs cost $137.50 for the month. Its 1,018 Slack replies cost $156. Less than half the runs, more of the money.
The model makes a big difference
| Model | Median run | 1 in 10 runs cost more than |
|---|---|---|
| Claude Haiku 4.5 | 1 cent | 1.4 cents |
| GPT-6 Luna | 2.7 cents | 4 cents |
| GPT-6.1 Sol | 4.1 cents | 14 cents |
| Claude Sonnet 5.5 | 7.4 cents | 19 cents |
| Claude Opus 5.5 | 25.7 cents | 43 cents |
The median Opus run cost about six times the median Sol run. Opus also scored highest on our own test set: 21 real tasks, 3 tries each, and Opus got 14 right all three times, Sol 10, Sonnet 9 and Haiku 8. The set is small and Opus did the grading, so treat that as a hint, not a ranking.
So we pick the model per job. My inbox agent runs on Luna and cost $1.73 for the whole month. My chief of staff runs on Sol, second on our tests at about a sixth of the price per run.
The bug that made one bill jump
In September one AI employee's cost per reply went from 2 to 4 cents up to 14 to 22 cents within two weeks.
That was our bug. AI companies give a big discount when the start of a prompt matches the last one (prompt caching). We were mixing text that changes on every message, like who is talking, into the part that should never change. So the discount almost never kicked in.
We moved the changing text out, and cache reuse on the same prompt went from zero to over 99%. We caught it within a day. Every job's cost shows in dollars, not tokens.
What you would pay
What you pay us depends on the job, and we tell you on the call, before anything starts. After that, every job shows what it cost in dollars, and you can set spending limits.
For a rough model-fee number on your own idea, use our October medians: 4.6 cents per scheduled job and 12 cents per Slack reply. 20 jobs and 30 replies a day comes to about $4.50 a day, or roughly $135 over 30 days.
Weighing this against building it yourself? Here's how WeaveKit compares with Zapier, n8n and Lindy.