Blog
What a unit of AI assistant work actually costs
Measured on Arlo in August 2026, a back-and-forth message costs about 8¢, a drafted document about 48¢, a researched answer with sources about 72¢, a built app about $2.64, a minute of live phone call about 11¢, and a minute of meeting notes about 2.5¢. The spread matters more than any single figure: a researched answer costs about nine times a message, which is what a flat per-seat price hides.
Almost every AI assistant is sold by the seat or by the month, which is a pricing model that answers the vendor's question (what is this account worth?) and not yours (what will the work cost me?). We meter Arlo per task, so we have the second number. Here it is.
These are our own measured rates as of August 2026, on our own system. They are not an industry benchmark, and another product built differently will land somewhere else. What we think travels beyond our pricing is the shape: the spread between task types is large, consistent, and almost entirely invisible under a flat monthly fee.
What one task costs
| Task | Cost |
|---|---|
| A back-and-forth message | ~$0.08 |
| A minute of meeting notes | ~$0.025 |
| A minute of live phone call | ~$0.11 |
| A generated image | ~$0.10 |
| A drafted email or document | ~$0.48 |
| A researched answer with sources | ~$0.72 |
| An app or dashboard, built | ~$2.64 |
The cheapest routine unit is a minute of meeting notes at about two and a half cents. The most expensive is a built app at $2.64 — roughly 105 times more. Even setting aside app-building as a special case, a researched answer costs about nine times what a message costs.
Why the spread is the whole story
A conversational turn is cheap because it is mostly one model call over context you already have. A researched answer is expensive because it fans out: searches, fetches, reading, discarding most of what it read, then writing with citations. Building an app is expensive for the same reason, several times over.
This is why "unlimited" and per-seat plans behave strangely in practice. If everyone on your team chats, the vendor's margin is enormous. If three people discover deep research and run it forty times a week, the economics invert — and the usual response is a rate limit, a slower model swapped in behind the scenes, or a quiet cap. You experience that as the product getting worse. You are actually watching a pricing model fail.
We have made both mistakes ourselves. An earlier version of our free tier metered real usage against a cap denominated in retail dollars rather than cost, which is the sort of error that only shows up once someone uses the product hard.
What this means when you are comparing tools
Three questions get you further than a price page:
- What does the work you actually do cost — not the average user's work? If your team's job is research, price the research. A tool that is cheap for chat and expensive for research is not cheap for you.
- What happens when you hit the ceiling? Ask specifically whether you get a hard stop, a slower model, or a surprise invoice. All three exist in the market; only one of them is honest, and it is the one that stops and tells you.
- Does unused capacity expire? Monthly allowances that reset punish uneven months, which is most months. Ours roll over for twelve.
The honest limits of these numbers
Per-task costs move. Model prices fall, context strategies change, and a re-measure can shift any row in that table — ours have moved before and will again, which is why every figure on this site is generated from one rate table rather than typed into the copy. Treat the ratios as more durable than the absolute cents.
And a caveat about the comparison: our costs are ours. We are publishing them because per-task numbers are almost impossible to find in this category, not because they settle which product is cheaper for you. If you want to check us, the pricing page derives every figure from the same table this post does.
For the work these numbers pay for, see AI research agent and AI operations agent. For where the category is heading on pricing generally, what is AEO covers a related measurement problem: what it costs to be found by the systems people now ask.
Frequently asked questions
How much does an AI assistant cost per task? On Arlo as of August 2026: about 8¢ for a back-and-forth message, 48¢ for a drafted document, 72¢ for a researched answer with sources, $2.64 for a built app, 11¢ per minute of live phone call, and 2.5¢ per minute of meeting notes. Other products priced per seat do not usually publish per-task figures.
Why do AI assistants cost so different amounts per task? Because the work is different. A chat turn is roughly one model call over existing context. A researched answer fans out into many searches and fetches, reads far more than it uses, and then writes with citations — so it costs about nine times a message on our system.
Is per-seat or usage-based pricing better for AI assistants? Per-seat is predictable and hides the variance; usage-based tracks what you actually consume. The risk with per-seat is what happens at the ceiling — rate limits, quieter model downgrades, or caps — because a flat fee has to protect its margin somehow.
What should I ask a vendor about AI pricing? What your specific workload costs rather than the average, what happens when you hit the limit, and whether unused capacity expires at the end of each month.
Last updated August 8, 2026