AI Agent Pricing Index

If you have tried to budget for an AI agent, you have probably found the same thing we did: a great deal of writing about pricing, and very little pricing. The “how much do AI agents cost in 2026” posts overwhelmingly cite each other.

So this is the boring version. We opened 38 vendors’ own pricing pages, wrote down every number printed on them, and recorded the date. Where a vendor publishes nothing, that is the entry. Nothing below is estimated, and nothing is taken from anyone else’s article.

The four things worth knowing

1. Roughly a quarter of this market will not tell you the price

10 of 38 vendors publish no usable paid price anywhere. Stack AI, CrewAI, Relevance AI, Voiceflow, Yellow.ai and Botpress show a free or entry tier and then a “custom” tier with nothing in between. Two go further — Sierra and Decagon have no pricing page at all; both /pricing URLs return a 404, and these are two of the best-funded companies in customer-service AI.

Two more are worth singling out because of what they choose to withhold. Gorgias states plainly that you “pay only when it resolves a conversation” and never says what a resolution costs — the same move Zendesk makes. Artisan publishes exactly how much work each tier does (~2,500 and ~6,000 leads contacted a month) and refuses to price either one. In both cases the billing unit is public and the number attached to it is not.

Widen it to vendors hiding any tier and it is 32 of 38. Published pricing in this market is an on-ramp, not a price list — the top of every ladder is negotiated.

2. The unit is usually the thing that gets you, not the number

We counted 24 distinct billing units across these pages. A credit, a task, an activity, a run, an action, an outcome, a resolution, a conversation, a session, a minute, a vCPU-hour, a browser-hour. Zapier bills two different units on one pricing page — tasks on the core product, “activities” on Agents.

This is why you will not find an average anywhere on this page. Averaging a per-resolution price against a per-vCPU-hour price produces a number with a decimal point and no meaning.

3. Prices that look comparable and are not

Vapi publishes $0.05/min and Retell publishes $0.07–$0.31/min, so Vapi looks cheaper. But Vapi’s page states model costs are “passed on to customer” and transport is “charged by provider, not Vapi” — that $0.05 is hosting alone. Retell’s range is all-in and itemised: infrastructure $0.055, speech $0.015, model $0.003–$0.16, telephony $0.015. One is a part; the other is the whole.

Bland AI shows the same shape a different way. Its cheapest per-minute rate, $0.11/min, carries a $499 monthly platform fee, while its most expensive, $0.14/min, carries none. The cheap one only becomes cheaper somewhere north of 16,000 minutes a month. Read the per-minute column alone and you would pick wrong.

The same trap sits inside Salesforce, which prints prices from $5 to $550 per user per month on one page — and attaches “Contact us” to all six of them.

4. The only two prices in this whole table that are genuinely like-for-like

AWS Bedrock AgentCore charges $0.0895 per vCPU-hour. Google’s Gemini Enterprise Agent Platform charges $0.085 per vCPU-hour. Same unit, same thing, within 5% of each other. Everywhere else in this dataset, comparison needs a footnote.

The full table

46 rows. Click any vendor to see every tier exactly as their page prints it, with a link to the page and the date we read it. Only 6 of 38 software vendors price every tier they offer.

What this costs against a person

The last rows are not software. They are there because the honest alternative to an agent is usually a person, and people have published rate cards too. Sorted by what a full-time equivalent costs per month:

• $35–$125 a month — Fancy Hands, sold as 3, 5 or 15 requests a month rather than hours at all
• $390–$2,160 a month — Time etc, for 10 to 60 hours ($39 down to $36 an hour). Sixty hours a month is the largest plan they publish
• $699 / $999 a month — Wing Assistant, for 80 or 160 hours
• $270 / $540 a week — Magic, part-time and full-time
• $1,600 per unit a month — Prialto, 55+ hours, plus a $250 set-up fee and a 90-day minimum
• $1,988 a month — MyOutDesk managed, rising to $2,500 specialised, for a full-time placement
• $3,000 a month — Athena, a dedicated full-time EA, monthly after a 90-day commitment
• $8–$16+ an hour — Virtual Latinos, by experience level, for full-time placement

Note that these are not directly comparable to each other either, for the same reason the software rows are not: some are hours, some are requests, some are a whole person, and only some include the recruiting and management. Where a vendor sells hours we have quoted the hours rather than converting to a full-time equivalent, because their largest published plan is often well short of full-time — Time etc tops out at sixty hours a month, not a hundred and seventy.

Our own price sits inside that range. $2,500/month. Month-to-month. Cancel anytime. At $2,500 a month you could instead have one MyOutDesk specialised placement, or two Wing Assistant full-timers, or most of an Athena EA. Those are different purchases and none is automatically right — a person handles the case nobody wrote down, and tells you when the process itself is wrong, which no agent does. What they are not is a like-for-like comparison, which is the point this whole page is making.

How this was put together

Each vendor’s public pricing page was retrieved directly and the figures copied verbatim. Three rules:

• If it is not printed, it is not here. A missing price is recorded as not published, never estimated and never taken from a third party’s claim about that vendor.
• If the page could not be read, the vendor is left out. Several vendors render pricing in JavaScript that we could not retrieve cleanly. Rather than guess, we dropped them. This index is smaller than it could be and correct.
• No averages. See point 2 above.

Prices change without notice. Every row carries the date it was read. If you find a row that is now wrong, it is wrong — tell us and we will fix it.

Use this on your own site

Free to embed, no attribution required beyond the link that ships with it.