n8n vs Zapier vs Make for AI Agents: Cost at Volume (2026)

A nine-step Zap that runs 1,500 times a month does not use 13,500 tasks. It uses 9,000, because Zapier does not count triggers, Filter steps or Formatter steps. The same workflow on n8n is 1,500 executions.

That gap, 9,000 against 1,500, is why every n8n vs Zapier argument turns heated. It is also why many comparison posts get the sums slightly wrong. This guide does the arithmetic from the vendors' own billing pages, then asks the harder question: when is none of the three the right answer?

How do Zapier, n8n and Make count usage?

Three platforms use three different meters. Mixing them up is the most common reason a price comparison is wrong.

Notice what the table does not say. Zapier does not bill every step. The shorthand that Zapier charges per step overstates the bill and makes n8n's advantage look bigger than it is.

Two rules catch people at invoice time. On Zapier's pricing page, pay-per-task overage runs at 2.5 times the base task rate on monthly self-serve plans and 1.25 times on annual ones. With pay-per-task off, usage simply stops when you reach your task limit. With it on, usage still stops at three times your plan. n8n's pricing page shows annual prices only, so a monthly-billing comparison needs a quote.

What does one real AI workflow cost on each platform?

Here is one workflow to price, the kind of lead flow sales teams ask for first. The step counts are an assumed example, not a measurement, and every lead is assumed to pass the filter.

1. Trigger: a new lead arrives from a web form. Free on Zapier.
2. Formatter: clean the name and company fields. Free on Zapier.
3. Action: look up company data from an enrichment API. One task.
4. Action: an AI model scores the lead from 0 to 100. One task, possibly more depending on the model tier.
5. Filter: continue only if the score is 70 or higher. Free on Zapier.
6. Action: create the contact in your CRM. One task.
7. Action: post an alert in Slack. One task.
8. Action: log the lead in a spreadsheet. One task.
9. Action: draft a follow-up email. One task.

Zapier counts 6 tasks per run. n8n counts 1 execution. Make counts about 6 to 7 credits, one per module run, with the range there because Make's help page does not say whether the trigger uses a credit.

At 100 leads a month, Zapier and n8n cost about the same, and Make's free plan covers the load. At 1,500 leads, Zapier's headline price is several times n8n's. At 10,000 leads, n8n's own price curve turns awkward, because its page lists no paid cloud tier between Pro's 10,000 executions and Business, which is €667 a month billed annually and, as of October 2026, self-hosted only.

Zapier lists plan tiers between the ones named above that were not itemised, so check the live pricing page before you commit. Make's own pricing page, read on 30 October 2026, names Core, Pro and Teams plans with 10,000-credit tiers but shows no paid prices in its text. Gamut's 8 October 2026 comparison lists Make's Core plan at $9 a month for 10,000 operations without saying whether that is annual or monthly billing, so confirm it on Make's page.

The model bill sits on top of all of this. If you call the same AI model with your own API key, you pay the provider the same tokens wherever the call is made. The exceptions are Zapier's AI by Zapier steps, whose task rate depends on the model tier, and Make's own AI Provider, which charges credits by tokens.

If you want to see this lead flow built end to end, the sales automation workflow guide walks through each stage.

What changes when the AI step is a real agent, not a single prompt?

A scoring prompt is one step. An agent is a loop: the model picks a tool, reads the result, then picks again.

By n8n's definition, one run of the workflow is one execution, so an agent that calls five tools inside the run is still one execution. Zapier Agents use activities instead: 400 a month on Free and 1,500 on Pro, capped at 10 activities per run on Free and 40 per run on paid plans. Make's agent chat uses credits based on tokens when you use its AI Provider.

That makes n8n the cheapest meter for loop-heavy agents. It does not make n8n the cheapest total, because model tokens are billed separately and grow with every loop. Set a cap on tool calls per run on any platform, since a runaway loop is how a small workflow turns into a surprise bill.

For the wider picture of what agents do in a business, read our guide to AI agents for business.

When should you leave Zapier, and where should you go?

Leaving Zapier is three different decisions, because three different problems cause it.

The row people skip is the first one. If the bill is the pain, the cheapest fix is almost never custom code.

Does price alone ever justify a custom build?

Rarely. Take the platform bill you would stop paying, subtract the running cost of the custom version, and divide the build cost by that saving. The figures below are illustrative arithmetic, not quotes.

In the 1,500-lead example, moving from Zapier to n8n saves roughly $100 a month, ignoring exchange rates. At that gap a custom build is nowhere near paying back on price.

Agency-published guides put a single-workflow agent at roughly $5,000 to $40,000 to build, with running costs of about $500 to $5,000 a month, which is directional rather than a quote. If custom running costs exceed your saving, the build never pays back at all.

CodeXcelerate's AI agent builds start from $2,000 and take 4 to 6 weeks. Price still is not the reason to choose one.

A decision rule you can reuse

Go custom when at least two of these are true:

1. The workflow is core to revenue, so an outage or a wrong answer costs real money.
2. It must read or write your own systems with logic no connector expresses.
3. You need tests, audit logs or approval pauses that live in version control.
4. After moving to n8n, the platform bill still exceeds the custom running cost plus the build cost spread over 12 months.

If fewer than two apply, stay on the platform or move to n8n or Make. The comparison of 13 AI agent platforms covers the wider field if none of these three fits.

What does LangGraph add that n8n and Make do not?

LangGraph is an open-source framework for building agents in code. A common claim says n8n cannot hold state, remember a conversation or wait for a long time, but n8n's documentation says otherwise. The real difference is where the state lives and who can test it.

In n8n, state is workflow data plus memory nodes. In LangGraph, state is saved as snapshots called checkpoints at each step, so you can replay what the agent knew when it made a decision. That matters when a wrong answer costs money.

It also comes with costs. LangGraph needs engineers, a database for checkpoints, since in-memory checkpoints are lost on restart, and a server to run on. n8n lets a technical operations person ship something useful in days.

CodeXcelerate builds agents on both LangGraph and n8n. Its AI Lead-Gen Machine case study runs on n8n with AI email and AI voice calls, which shows that an AI outreach system does not always need a from-scratch codebase.

Does location change the bill for US, UK and Australian teams?

It changes the currency first. Zapier prices in US dollars and converts at the exchange rate on the payment date, while n8n lists its Cloud plans in euros. A UK or Australian team therefore compares two prices that both move with the exchange rate.

Australia adds a legal point from 10 December 2026. The OAIC says organisations covered by the Australian Privacy Principles (APP entities) must then describe in their privacy policy the kinds of personal information used by computer programs that make, or substantially drive, decisions that could significantly affect someone. The OAIC's notice refers to computer programs, not only AI, so a no-code scenario that approves or rejects people can sit in scope as readily as an agent.

This is general information, not legal advice. Our guide to Australia's new automated decision-making rules shows what to check in your own workflows. If you hire a team in India to build or migrate automations, CodeXcelerate keeps 3 to 4 hours of daily real-time overlap with US and UK time zones.

How do you test a move before you switch?

Do not rebuild everything at once. Prove the saving on one workflow first.

1. Count real tasks. Export last month's Zap history and count successful action steps per run.
2. Do the sums. Tasks equal runs times counted actions. Executions equal runs.
3. Rebuild the busiest workflow on n8n or Make with the same trigger.
4. Run both side by side for about two weeks, with the new version writing to a test sheet.
5. Compare outputs and failures, then fix the gaps.
6. Switch the live writes, keep the Zap paused for a month, then cancel or downgrade.

Use this worksheet to price any workflow before you decide.

WORKFLOW COST WORKSHEET
 
Workflow name: ______________________
Runs per month: ______
 
ZAPIER
Counted actions per run (skip the trigger, Filter, Paths, Formatter, Delay): ____
Tasks per month = runs x counted actions = ______
Plan tier that covers it (check zapier.com/pricing): ______
 
N8N
Executions per month = runs = ______
Sub-workflow runs do not use up the monthly limit.
Cloud plan that covers it (check n8n.io/pricing): ______
 
MAKE
Modules per run (check whether the trigger counts): ____
Credits per month = runs x modules = ______
 
ALL PLATFORMS
Model tokens per run x runs x provider price = ______
Maximum tool calls allowed per run (set a cap): ____
What does one wrong answer cost? ______

That last line is the one most spreadsheets leave out. It decides whether you stay on a platform or write code.

The number that matters is not the plan price on the pricing page. It is runs multiplied by counted steps, run through each platform's own meter. Do that sum before you rebuild anything, because it often shows that the cheap fix is moving to n8n or Make, not hiring an engineer.

Custom code solves a different problem: risk, control and state that you can test. If you cannot tell which problem you have, send CodeXcelerate the workflow and its monthly run count, and the team will tell you honestly whether to stay, move to n8n or build custom through the AI agents service or a free 30-minute call.