AI agent pricing comparison: what you'll actually pay

What you're actually comparing

Type "AI agent pricing comparison" into Google and you'll get three products wearing the same label. A meeting note-taker on a monthly subscription. An automation platform billing per API call. A custom-built agent that a consultancy scoped and priced as a project. They solve different problems, so putting their sticker prices side by side tells you little unless you know which bucket you're shopping in.

Circleback, the meeting note-taker, just gave us a clean, dated example of how the subscription side works: what's free, what's gated, and what the free tier is actually for. We'll use it, then walk through the other two models, so you can tell which one you're actually being sold.

Subscription pricing: Circleback and the rest of the pack

Circleback announced on 31 August 2026 that it's adding a free tier, its first ever. Before this, every plan started at $20.83 a month. Now you can record meetings, transcribe an unlimited number of them, query the transcripts with AI, and connect the app to Linear and Slack, all for nothing. The catch is your meeting history: the free plan only keeps the last 30 days. If you want every integration, unlimited history, and full API and MCP access, the paid plans start at $15 a month, billed annually.

That's the subscription model in one example. The free tier is customer acquisition. Founder Ali Haghani said the company doesn't spend on Google or Meta ads at all, the free tier is the marketing budget. Circleback was founded in 2023, raised $2.5 million in 2024, and has been profitable since, running annual recurring revenue of roughly $8 million with a team of eight.

The pattern holds across most subscription AI agents, whether it's a note-taker, a calendar assistant, or a sales copilot. The free or entry tier gets you using the product, and the paywall sits on the things that make it sticky: integrations, history, and API access, rather than the core AI feature itself. Granola, a competitor in the same space, added a similar 30-day history limit on its free tier a few months earlier.

Usage-based pricing: paying for what the agent does

A second group of AI agents don't charge per seat at all. They charge for what the agent actually does: a fee per API call, per token processed, or per workflow execution. This is how most automation platforms and custom agents built on a model provider's API get billed underneath, even when the platform on top adds its own markup or subscription layer.

The appeal is that cost tracks usage. An agent that processes five invoices a week costs you almost nothing, one that processes five thousand costs proportionally more. The downside is the same thing: nobody can hand you one number for what an AI agent costs, because the answer depends on how much work you point it at. If you're comparing vendors on this model, ask for the price per unit of work, not a headline monthly figure, and run your own volume through it before you sign anything.

Custom agents: priced as a project

The third model is the one we work in most: a custom-built agent, priced as a project rather than a subscription. A business doesn't want a generic meeting note-taker, it wants something that reads incoming purchase orders against contract terms, flags the ones that don't add up, and drops the rest into the existing approval flow. Nobody sells that off the shelf, because it's wired into your systems and your rules.

Here the price tag is a build fee for the scoping and integration work, plus the ongoing, usually modest, cost of the underlying model calls once it's running. The "seat" is the workflow rather than a person logging into a dashboard, so there's no per-seat charge at all. This is the kind of work we scope through our AI services: we sit with the team doing the task now, work out what the agent actually needs to see and decide, and price the build against that, not against a generic subscription tier.

Doing an AI agent pricing comparison across three models

Line the three up and the differences are less about the number and more about what the number is measuring.

Model	Typical use	What you're paying for	Example
Subscription / per seat	Note-takers, calendar agents, sales copilots	Seats, integrations, history depth	Circleback: free with 30-day history, $15/month billed annually for unlimited history and API access
Usage-based	Automation platforms, API-driven agents	Tokens, calls, or executions	Cost scales directly with volume processed
Project fee plus hosting	Custom workflow agents	Scoping, integration work, ongoing model cost	Built around one specific task inside your own systems

A solo founder taking client calls wants the subscription tier. They're not going to out-build a $15-a-month plan on their own. A business running a workflow nobody sells as a product needs the third column, and paying for a subscription tool that almost fits is often the more expensive choice, because "almost fits" means someone still checks its work by hand every week.

What to ask a vendor before you sign

Ask what happens to the price at ten times your current volume, not just at today's. Ask whether the feature you actually need, API access, a specific integration, unlimited history, sits behind a plan tier you weren't planning to buy. Ask whether you're being billed per seat for a tool that only one person on your team will ever touch, which happens more often than vendors like to admit.

Circleback's move is a useful signal: a profitable, self-funded company chose to give the core product away and charge only for what locks you in, unlimited history, every integration, API and MCP access. If a vendor's free tier looks generous, check which of those three it's still charging for.