n8n vs Zapier vs Make: what we actually deploy and when

We’re an automation agency that self-hosts n8n, so you already know where this comparison lands. Here’s the part that should earn it some trust anyway: sometimes the answer we give after an audit is “use Zapier — here are the four zaps, you don’t need us for this.” The right platform is a function of your volume, your team, and your stack, not our preferences. This is the decision logic we actually use.

One caveat before the table: list prices below are the shapes as of mid-2026. Vendors move numbers around, so check their pricing pages before deciding on cost alone. The pricing models, though, have been stable for years — and the models are what decide the outcome.

The comparison, without the fog

Why we default to self-hosted n8n

Self-hosted n8n means no per-task fees: 100,000 executions a month for the price of a VPS.

• ▸The cost curve is flat. Per-task pricing means the bill grows with your success: automate more, pay more, forever. And the metering units differ — an n8n execution is a whole workflow run, so a 40-step workflow is 1 execution. The same logic in Zapier bills ~39 tasks per run.
• ▸Code is available, not forbidden. Half of real business automation is the unglamorous 10%: parsing a malformed CSV, retrying a flaky API, signing a custom auth header. In n8n that’s a JavaScript or Python node in the canvas. On the other two, it’s the wall you hit.
• ▸Failure is designed for. Every workflow we ship carries error workflows, retry logic, and alerting — automations that fail loudly, not silently. n8n makes that a first-class pattern rather than a paid add-on.
• ▸Your data stays yours. Invoices, payroll, patient records — self-hosted means nothing transits a third party’s cloud. For regulated or GDPR-adjacent work this isn’t a preference; it’s the requirement.
• ▸Workflows export as JSON. Version control, code review, staging copies, rollback — the automation is software and gets treated like it.

The honest downsides

• ▸Somebody has to run the server: updates, backups, monitoring. That’s either your technical person or a care plan (ours start at $500/mo). If it’s neither, keep reading — Zapier exists for you.
• ▸Fewer polished native integrations. The HTTP node covers nearly anything with an API, but “nearly anything” takes an hour of building where Zapier takes a minute of clicking.
• ▸The learning curve is real. n8n is a developer-shaped tool wearing a visual interface. Your ops manager can run and read the workflows; they probably shouldn’t build them from scratch on day one.
• ▸License fine print: the free Community Edition ships under n8n’s Sustainable Use License — internal business use is explicitly allowed, which covers running your own automations. Reselling hosted n8n itself is not. (n8n Cloud exists from about $25/mo if you want n8n without the server.)

When we recommend Zapier anyway

This is the section most agency comparisons skip, because it costs them projects. Below certain thresholds Zapier simply wins, and pretending otherwise would tell you exactly how much to trust the rest of the article:

• ▸Under ~1,000 tasks a month. You’d pay Zapier $20–50/mo. A VPS plus anyone’s maintenance time costs more than that. The math isn’t close, and the math is the argument.
• ▸Nobody technical owns it. Zapier’s entire product is “nobody maintains this.” If your team is 4 people and none of them wants to think about a server — ever — that’s worth real money every month.
• ▸Your apps are Zapier-first. With 8,000+ integrations, the long tail of niche SaaS often ships a polished Zapier integration and nothing else. If your stack lives in that tail, fighting it is a bad trade.
• ▸Your process changes weekly. Zapier’s editor is the fastest iteration loop in the category. For a workflow still finding its shape, that speed beats architecture.

The crossover, typical for this type of decision: somewhere between 2,000 and 10,000 tasks a month — or the first day a workflow needs a loop, a queue, or a code step. That’s when the per-task meter and the linear editor start costing more than a server would.

When Make wins

• ▸Mid-volume on a small budget. Roughly 10,000 operations for about $10/mo is the cheapest metered automation on the market.
• ▸Visual data surgery. Make’s iterators, aggregators, and array handling genuinely beat Zapier’s for reshaping data — closer to n8n’s power, without the server.
• ▸Debugging by eye. Watching data flow through a Make scenario, module by module, is the best non-developer debugging view of the three.

Where Make loses us: no self-hosting, and operations metering that punishes chatty workflows — every module run bills, so a 10-module scenario at 1,000 runs a month is 10,000 operations, not 1,000. And its code limits arrive exactly when the workflow gets interesting.

The rule we use

A worked example, with the meters running

Take one concrete workflow: lead intake. Form submission → enrich → score → CRM row → Slack ping → follow-up email. Six steps, and say it fires 5,000 times a month. On Zapier that’s ~25,000 billed tasks (the trigger is free, the 5 actions aren’t) — a mid-tier plan in the low hundreds per month at list. On Make, ~30,000 operations — roughly $30–60/mo. On self-hosted n8n, 5,000 executions — the same $20 VPS you’d run at 500 or at 50,000. Approximate figures, list prices drift; the shape of the curves doesn’t.

What this means for a build

Our build pricing doesn’t change with the platform — a workflow build is $2k–6k fixed whichever tool fits, because the work is the process design, the error handling, and the integration depth, not the logo on the canvas. What changes is what you pay after we leave: a Zapier deployment carries its per-task meter forever, while a self-hosted n8n deployment runs on tool fees only — typically $40–180/mo all-in. Over a 3-year life, that difference is usually larger than the build cost itself.

So the platform question is really a total-cost question wearing a features costume. Count your monthly runs, name who’ll own the system, list the 6 apps that matter — the answer usually falls out on its own.