n8n vs Zapier vs Make: what it costs after 90 days of real workflows

After 90 days the invoice is hosting, failed runs, and the person who owns the scenario — not the starter price. Zapier bills tasks, Make bills credits, n8n bills executions. Price one inquiry workflow on the official pages, then add the first agent webhook. The cheap plan is rarely the cheap system.

The useful comparison is not week one. Week one is a trial and a Slack ping. Day 91 is a webhook the agent posts to, a CRM that changed a property name, three failed runs nobody saw until Monday, and a person who is now “the Zapier person” whether that was the job or not.

We ship n8n workflows the team can own. We also inherit Zapier and Make accounts. This is the arithmetic we do before we touch either.

What do the official comparison pages actually claim?

We opened the three vendor pages.

n8n’s Zapier comparison is explicit: n8n charges per full workflow execution; Zapier charges each task. n8n’s line is the one they want remembered: you should never be penalized for building complex workflows. Webhooks sit on every n8n plan. Zapier treats Webhooks as premium from Pro up.

Zapier’s 2026 n8n comparison (updated June 2026) answers from the other side. Ryan Kane writes that broken workflows at 2 a.m. and accidentally exposed credentials are not his idea of a good time. They treat n8n as self-hosted infrastructure; Zapier as a managed suite. Their pricing claim is the inverse: Zapier’s price is the visible one; n8n’s real costs are hosting, maintenance, and the engineer.

Make’s positioning is the visual middle. Credits replace operations. AI agents and Make Grid sit on paid plans. The 2026 compare table lists Core at $9/month for 10,000 credits, against Zapier Professional at $19.99 for 750 tasks.

All three pages are sales documents. Read them as claims. Then price a workflow you actually run. None of them invoice the person who gets the error email.

How do n8n, Zapier, and Make bill a real workflow?

The unit decides.

Figures above are from the public pages on 13 August 2026. Zapier’s pricing reference: triggers, Formatter, Paths, Filter, Delay, and Looping do not count; a successful action does. Overage is 1.25× on annual billing or 2.5× monthly. Make moved from operations to credits on 27 August 2025 (n8n vs Make); standard modules convert one-to-one, AI and Code do not. Treat $9 as the annual Core sticker, and open the calculator before you commit.

A two-step “new row → Slack” makes Zapier look cheap. An eight-step inquiry desk does not.

What is a client-shaped scenario we can price three ways?

We do not price a template. We price the desk: an inquiry arrives, a record is written, a person is told, a Monday board can be read. Same spine as the Aurelia support desk — ticket in, a person only when needed — minus inventing their volumes.

The Monday inquiry desk (one scenario, three bills)

1. Webhook receives a JSON payload from the site form or the agent.
2. Validate: required fields present, email looks like an email.
3. Lookup the contact in the CRM.
4. Create or update the contact.
5. Open a ticket or a deal.
6. Post a line in Slack.
7. Append a row to the sheet the weekly board reads.
8. Send the confirmation email.
9. On failure: notify the owner, do not silently drop the inquiry.

Volume used below is an estimate, not a client metric:

• 400 form inquiries / month
• 200 agent handoffs / month
• ~50 replays and tests that hit production
• 650 production runs / month
• 90 days ≈ 1,950 runs

Zapier billable steps per successful run: six (CRM lookup, write, ticket, Slack, sheet, email). Filter and Paths are free. The Webhooks app requires Professional. Make: about eight modules plus an error route. n8n: one execution per run; a dedicated error workflow is a second execution when it fires.

Minimum that lets the campaign and the agent share a page without a human copying the row.

What does that scenario cost after 90 days?

Platform fees only. Annual billing unless noted. Public pages, August 2026; anything we multiplied is marked estimate.

Zapier — tasks

Estimate: 650 runs × 6 billable actions = 3,900 tasks / month, if every run succeeds and nothing extra fires.

That sits under Professional 5,000 at $89/month annual, or $133.50 monthly, per Zapier’s tier table.

Failed actions do not consume tasks. A successful retry does. A loop that writes 40 sheet rows is 40 tasks. An “AI by Zapier” classify step is not one task — see the rates page. We are not inventing that number.

Make — credits

Estimate: 8 modules × 650 runs = 5,200 credits / month if every module is 1 credit and nothing bundles.

That fits Core at 10,000 credits. Make’s compare page: $9/month annual. Monthly is higher on the live calculator — publicly cited around $12 at the 10k mark. Confirm before you buy.

A busy week that iterates 200 extra contacts is not 200 extra scenario runs. It is 200 extra credits per module that touches those bundles. Data transfer on paid plans is 5 GB per 10,000 credits (Make’s table). File-heavy work hits the ceiling first.

n8n Cloud — executions

Estimate: 650 inquiry runs + 40 error-workflow runs + 90 digest runs = 780 executions / month. Starter includes 2,500 at €20/month billed annually.

A 40-node workflow and a 4-node workflow cost the same. That is the execution advantage n8n sells.

The 90-day platform bill, side by side

The sticker is not the cost. The next three sections are.

What does hosting actually add?

Zapier and Make include hosting.

n8n Community Edition is free on GitHub. Free is the licence, not the machine. DigitalOcean Droplets start at $4/month (512 MiB) and $6/month (1 GiB). A production webhook does not belong on 512 MiB. Weekly backups add 20% (DigitalOcean).

n8n Cloud hosts on Azure in Frankfurt. If EU residency is why you looked at n8n, Cloud already answers. Self-host because you need the box or the volume — not because $6 looks like a win against €20.

n8n Business at €667/month is self-hosted SSO, Git, environments. Zapier Team at $69 is not the counterpart. Zapier Enterprise is. Do not drop Business into a starter table and call n8n expensive.

What do failed runs cost that the pricing page hides?

Failed work is where the three models stop being comparable.

Zapier’s public pricing reference is blunt: a task is counted when Zapier successfully completes a unit of work. Failed actions are not counted. Autoreplay (paid) retries temporary errors. You still lost the inquiry. You did not also pay for the miss.

Make charges the modules that ran. An HTTP module that gets a 500 still did work. An error-handler route is more modules. Overage protection sits on higher plans. Core does not promise you a free accident.

n8n Insights reports total production executions and failed production executions separately. Production means published workflows fired by a schedule or a webhook, including error workflows. Manual tests and sub-workflows do not count toward the parent. A failed production run still occupies the execution quota. A tidy error workflow is a second execution.

The cost that does not appear on any invoice is the inquiry that vanished. We count inquiries, not successful node checkmarks. If HubSpot rejects a property and twelve forms disappear, the platform bill is not the story.

Plan for failure the way you plan for volume: one error path that notifies a person; the same webhook twice must not open two tickets; a replay that does not double-write; a log Monday can read. On Aurelia, n8n opens the case and tags the product. The failure mode is “pull a person in,” not “retry until the quota dies.”

Who owns the scenario, and what does that person cost?

Someone owns it. If you cannot name them, it is already in trouble.

That person rotates the HubSpot token, notices the sheet is full, and decides whether the agent may write deal_stage. They read the 01

error at 07. None of the three vendors invoice this.

Zapier’s argument is that the owner can be anyone. Copilot drafts the Zap; Team shares connections; the trade is the task bill. n8n’s argument is that the owner should be technical — expressions, Code nodes, Docker. Zapier quotes a Reddit developer: most of their workflows have code nodes. That matches what we see. Make sits between: the canvas is readable in a meeting; the functions are Make’s own language.

We do not publish a fake loaded salary. We do publish the hours we actually see after month one, as a studio estimate:

Month one is two or three times that. A HubSpot property rename is an afternoon on every stack.

Price the owner before you celebrate the $9 plan. Two hours a week at whatever you already pay that person will dwarf Make Core and n8n Starter. It will not always dwarf Zapier at 50,000 tasks. TCO flips with volume, not brand preference. Name the owner on the workflow.

What happens when you add the first AI-agent webhook?

This is the join. The campaign makes the inquiry. The agent finishes the conversation. Something has to write the record. The first webhook is a POST: identity, intent, what was promised, what a person must review. Not a chat log dumped into Slack.

On n8n the Webhook node is a trigger with a test URL and a production URL. Activating the workflow is what makes production live. Teams forget that and debug against a cold test URL.

Zapier will not receive that POST on Free. Make will, with two active scenarios and a 5-minute max run — fine for a proof, not a desk.

The payload is the project. If the agent sends prose, the scenario guesses. If it sends fields — email, intent, sku, handoff_reason — the scenario files. That is the document being true, applied to a webhook.

YourGPT belongs here because that is the agent we connect. It answers from the client’s writing. n8n files the ticket. A chatbot that only talks never hits this section. An agent that takes action does. Ecommerce teams comparing builders should read the Shopify-adjacent agent note before they add a second brain in Zapier Agents. Two agents writing the same ticket is a race.

Which tool should you pick after 90 days?

Pick from the desk, not the landing page.

Stay on Zapier when the team is not technical, the app only exists there, and the task graph is short. Pay Professional. Turn Autoreplay on.

Move the heavy desk to Make when the canvas matters and credits beat tasks at your step count. Watch bundles in week two.

Put the desk on n8n Cloud when the workflow is long and the agent webhook is real. Starter at €20 covers the volumes above.

Self-host n8n when data cannot leave the VPC and a named person will patch the box. Otherwise you have bought a hobby.

We default to n8n for new work because the agent, the site, and the ads should share one log. We do not migrate a quiet two-step Zap to look principled.

The 90-day number is not “n8n is cheaper.” It is platform + hosting + failed-run policy + named owner. If that sum is not on one line, you have a preference, not a comparison.

FAQ

Is n8n cheaper than Zapier after 90 days?

Often on long workflows, not always on short ones. Official Starter is €20/month for 2,500 executions; official Professional starts at $19.99/month for 750 tasks. Our desk estimate is 3,900 Zapier tasks ($89/month at 5,000) versus 780 n8n executions (€20). Add a named owner. A two-step Slack ping can be cheaper on Zapier. Price your step count.

Does Make include AI agents in the base plan?

Yes on paid plans, according to Make’s 2026 compare pages. Make AI Agents and Make Grid are listed in the paid packages; Zapier sells Agents and Chatbots as add-ons with their own allowances. “Included” still means credits. An agent that calls tools will spend them. Confirm current packaging on make.com/en/pricing before you brief a client.

Do failed Zapier runs use tasks?

No. Zapier’s public pricing reference says failed actions are not counted. Successful retries are. That is a real advantage when an API is flaky. It is not a substitute for an error path. The inquiry can still disappear; you simply do not pay a task for the miss. Build the notify-a-person path anyway.

Can I self-host Zapier or Make?

No. Zapier is cloud only. Make is cloud only; Enterprise can install an on-prem agent for local systems, and scenarios still execute in Make’s cloud. n8n is the one of the three you can run on your own machine, from Community on GitHub up through licensed Business and Enterprise.

What should the first agent webhook contain?

A short JSON contract: who, what they asked, what the agent promised, what a person must review, and an id you can replay. Not the full transcript in the CRM description field. The scenario should validate, write once, and tell a person when it cannot. If you cannot describe the payload on one slide, you are not ready to turn the agent on.

When is n8n Business worth €667 a month?

When you need self-hosted collaboration, SSO, Git, and environments, and Community is no longer tenable. It is not the next step after Starter for a single inquiry desk. Starter and Pro are the Cloud path. Business is the licensed box. Compare it to Zapier or Make Enterprise, not to $19.99 Professional.

Should we migrate existing Zaps to n8n?

Only the ones that hurt. A quiet two-step Zap is not a migration. A multi-step desk that is 80% of the task bill is. Rebuild, do not expect an importer to save you. Keep the Zap live until the n8n webhook has replayed a week of real payloads without double-writing.

Send the current Zapier or Make invoice and a list of the live workflows — names, step counts, and last month’s task or credit use. We will write back with what that same desk would cost on n8n Cloud and on a Droplet, and which scenarios we would not move. The contact form is enough; a CSV export is better.