Every price below was read from the vendor's own page on 13 November 2026. Where a page did not show a price, we say so rather than guess. This is not a ranking: the setups are the jobs people most often hand to agents, and the tools are examples with public prices.

How agent bills work

Almost every agent bill has three layers.

  • The plan. A monthly or per-seat fee for the tool itself.
  • Usage. Tokens sent to a language model, minutes of audio, scraping credits, contacts researched, API calls. This is the layer that grows when the agent gets busy.
  • Review time. Somebody still reads the summaries, approves the emails and fixes the odd mistake. It never appears on an invoice, and it is often the largest cost.

Tools differ mainly in which layer they put the price on. A flat plan is easy to budget but wasteful when idle; pure usage pricing is cheap at low volume and can run away at high volume.

The model layer

If you build your own agent, the language model is usually the biggest usage line. Anthropic's API prices per million tokens (MTok) on 13 November were:

  • Claude Haiku 5.5: $0.10 input, $0.50 output (prompts up to 100,000 tokens)
  • Claude Sonnet 5.5: $2 input, $10 output
  • Claude Opus 5.5: $4 input, $20 output
  • Claude Fable 5.2: $8 input, $40 output

Batch processing halves those prices for work that can wait, and cached prompt reads cost far less than fresh input ($0.20 per MTok on Opus 5.5 against $4).

A worked example, with an assumed workload: an agent task that reads 20,000 tokens and writes 2,000, run 1,000 times a month. That is 20 million input tokens and 2 million output tokens. On Haiku 5.5 it comes to about $3 a month; on Sonnet 5.5, $60; on Fable 5.2, $240. Same job, an 80-fold spread, which is why choosing the smallest model that does the job well is the first cost decision, not the last. Our Claude profile covers the consumer plans.

Eight setups, itemised

1. Meeting notes and follow-ups

The off-the-shelf route is Fireflies: free with limited AI summaries, Pro at $10 per seat a month billed annually ($18 monthly), Business at $19 ($29 monthly). Fireflies widened its assistant, Fred, into email, Slack and search on 11 November, under the same subscription.

Building it yourself means paying for transcription and a model. AssemblyAI lists pre-recorded transcription at $0.15 an hour (Universal-2) or $0.21 (Universal-3.5 Pro). Deepgram lists streaming at $0.46 an hour on pay as you go. Forty hours of recordings a month is $6 to $8.40 at AssemblyAI's rates, before the summarising model. For one team, the subscription is simpler; the per-hour route wins when you are processing thousands of hours.

2. Workflow automation with an AI step

n8n charges by workflow execution: Starter is €20 a month billed annually for 2,500 executions, Pro €50 for 10,000, and the self-hosted community edition is free. Zapier has a free tier of 100 tasks a month, Professional from $19.99 and Team from $69. Make blocked our price check, so see its own page. In all three, a step that calls a language model with your own API key adds the model bill on top.

3. A coding agent on a subscription

Claude Pro is $17 a month with annual billing ($20 monthly); Max is $100 a month for five times Pro's usage and now includes $100 a month of API credit. Tools that ride on that plan add little: Rill, a Mac browser that hands pages to Claude Code or Codex, is free, and Faylee bug reports are free for two members or $49 a month for a team of ten.

4. Outbound sales

AiSDR's Ami charges pay as you go from $1.50 per prospect, or $250 a month for the Solo plan with 200 researched contacts. At 200 contacts a month the plan beats pay as you go ($300); at 166 contacts or fewer, pay as you go is cheaper. Pancake, which gives your own agent data and sending tools, starts at $99 a month for 60,000 credits.

5. Web research and scraping

Context.dev gives 1,000 free API credits a month (about 1,000 standard scrapes), then $20 a month for Developer and $90 for Pro. Manifold, for marketing data, gives 250 free credits and then $29, $79 or $249 a month.

6. A phone agent

Deepgram's Voice Agent API lists $0.075 a minute on pay as you go for its standard tier. With an assumed 1,000 calls of three minutes each, that is 3,000 minutes, or $225 a month. Vaani, priced in rupees, charges a platform fee of ₹1 a minute plus ₹600 a month for a phone number.

7. An agent that buys its own tools

Monid gives an agent one prepaid balance for more than 2,800 paid APIs, charged per call (its home page shows $0.0013 for one example call) with $1 of free credit to start. No subscription, but also no ceiling unless you set one.

8. AI in a shared team workspace

Hops charges nothing per seat and sells credits at $20 per 20,000. Its own guide suggests budgeting $20 per active user a month and says a deep research report uses 500 to 1,500 credits, roughly a dollar.

Where bills surprise people

Context sent again and again. An agent re-sends its instructions and history on every step. Our Hermes cost breakdown traced one bill to a 14,000-token preamble sent hundreds of times a night. Prompt caching is the fix where the provider supports it.

Unattended loops. Anything billed per call, per credit or per minute can be spent by an agent stuck in a retry loop. Set hard caps: a small wallet balance, a monthly spend limit on the API key, a maximum number of steps per task.

Seats versus usage. Per-seat plans punish occasional users; usage plans punish heavy ones. Count who will actually use the tool each week before choosing.

Working out whether it pays

The useful number is cost per completed task, not cost per month. Take one week of real work, count the tasks the agent finished without rework, divide the week's bill by that count, and compare it with what the same task costs you in time. Do the review-time sum honestly: an agent that saves ten minutes and needs eight minutes of checking is a two-minute agent.

If you plan to sell agent work to clients, our Build an AI Agency course covers pricing it, and the AI finance tools directory lists tools for tracking the spend.

Sources