Five platforms, five different units
Every figure is as of 12 September 2026, taken from each vendor's own price list.
| Platform | Name of the unit | What it actually measures |
|---|---|---|
| Zapier | task | every successful action step |
| Make | credit | every module operation (1 operation = 1 credit by default) |
| n8n | workflow execution | one complete run — regardless of the number of steps |
| Pipedream | credit | 30 seconds of runtime on 256 MB of memory |
| Power Automate | user / bot | a per-head quota, not consumption |
The third row is the important one. n8n's own wording, literally: a single execution is one run of your complete workflow — it does not matter how many steps it has or how much data it processes, it is still one execution.
What does that mean in money?
Scenario: 30 runs a day, so 900 a month.
| Steps / run | Zapier tasks | Make credits | n8n executions |
|---|---|---|---|
| 1 | 900 | 900 | 900 |
| 3 | 2,700 | 2,700 | 900 |
| 5 | 4,500 | 4,500 | 900 |
| 10 | 9,000 | 9,000 | 900 |
| 20 | 18,000 | 18,000 | 900 |
(Own calculation. On Zapier the trigger does not count, on Make it does; that makes no material difference in this table.)
And the same thing against the included quotas:
- n8n Starter (€20/month, 2,500 executions): the 900 runs fit comfortably even at twenty steps. In fact at any number of steps.
- Make Core ($12/month, 10,000 credits): the quota fills up at eleven steps — with the trigger that makes twelve modules, 10,800 credits. At twenty steps it is 18,000 credits — close to double the allowance.
- Zapier Professional (from $19.99): twenty steps means 18,000 tasks a month, for which you already have to buy a higher task tier, not the entry price.
The decision rule
It follows directly from this, and it is the point of the whole article:
| If your work looks like this | Then this favours you |
|---|---|
| Few steps, many runs (e.g. 5,000 runs × 2 steps) | Zapier / Make — the step-based model is competitive here, while n8n's execution-based quota fills up fast |
| Many steps, fewer runs (900 runs × 20 steps) | n8n — dramatically cheaper, by a factor of twenty |
| Long-running, compute-heavy work | Pipedream's time-based model penalises you — but with many fast steps the whole thing fits into a single credit |
| Many users, unpredictable load | Power Automate's per-head quota is expensive, but independent of consumption |
So the practical question is not "which is the best", but whether your workflow is the long kind or the frequent kind.
(The table compares consumption, not the final bill. Every platform behaves differently when the quota runs out — a tier change, an overage charge or a stop — and not every vendor publishes exact overage prices.)
Today's prices
| Platform | Entry paid plan | Quota |
|---|---|---|
| Zapier Professional | $19.99/month | selectable task tier, from 750 |
| Make Core | $12/month | from 10,000 credits |
| n8n Starter | €20/month (annual: €240) | 2,500 executions/month, 5 concurrent |
| Power Automate Premium | $15 / user / month | per-head quota |
| Power Automate Process | $150 / bot / month | unattended automation |
Pipedream's list price is deliberately missing from the table: its pricing page is built from JavaScript, and we could not confirm its plan prices from a primary source — only the definition of the credit. Free tiers: Zapier 100 tasks/month, Make 1,000 credits/month. One important and rarely mentioned limit: on Make's free plan the shortest polling interval is 15 minutes, on paid plans 1 minute.
Every n8n plan comes with unlimited users and unlimited workflows — against a per-head model that is a serious difference.
What does not count towards the quota
Worth knowing, because this is what upends your estimate.
By Zapier's own definition a task is every successful action step — and only the successful ones count. What does not count: no trigger at all, filters and paths, every step that errors out or halts, and several built-in helper steps.
Two consequences. The good one: checking frequently is free, because the trigger consumes nothing. The less good one: a replay pays again — if you rerun an execution, the steps that already succeeded eat tasks a second time. That is the hidden cost few people talk about.
On Make the base rule is "one operation, one credit", but there are exceptions: some compute-heavy modules consume two, ten or a variable amount, and AI modules consume dynamically, by token. Error handlers and paths, on the other hand, consume nothing.
Self-hosting: n8n is not open source
This is the most common mistake about the topic, and it is worth knowing precisely.
n8n's licence is the Sustainable Use License, and its decisive sentence reads: you may use or modify the software only for your own internal business purposes or for non-commercial and personal use; and you may only pass it on if you do so free of charge and for non-commercial purposes.
In plain terms, translated for a business owner:
- Yes: you may run it for free inside your own company, on your own processes — as a commercial company too.
- No: you may not pass it on as a paid service ("I host n8n for my clients").
- No: you may not build a paid product on it that is itself n8n delivered to someone.
This model has a name of its own — fair-code — and the definition itself states that the software is commercially restricted by the authors. This is not open source, and the fair-code site also defines itself as an alternative model. Several other well-known projects run under a similar licence.
And a warning worth knowing in advance: there are features you do not get even self-hosted without a paid licence key — custom variables, environments, external secrets management, log streaming, projects, SSO, workflow and credential sharing, and Git version control.
In other words: perfect alone, not in a team. Sharing, permission management and version control — exactly what a team needs — are paid.
(There is also a new, free intermediate tier: the email-registration "Registered Community" level, which adds folders, in-editor debugging and custom execution data. Many earlier comparisons do not include it yet.)
Of the others, the cloud versions of Zapier, Make and Power Automate cannot be self-hosted.
AI nodes and the model cost
The key question: is the model cost included in the platform fee? By default, no.
n8n: it can be connected to several model providers, and several agent types are available as separate nodes — including the classic ReAct pattern. There is vector store integration, there is human approval ("require human approval before an AI Agent executes specific tools"), and there is an evaluation feature. By default you pay the model cost with your own API key, directly to the provider.
A new option arrived in 2026, and its conditions are worth knowing: n8n Gateway credits is a prepaid balance that lets you use models and tools without an API key. But: it is only available on the cloud Starter and Pro plans — not on Enterprise and not self-hosted — unused balance is not refunded, it expires after 12 months, and it is lost if the account is closed.
Make: the AI agents are in beta and available on every plan, with your own key or through the platform's provider. The AI modules, however, consume credits dynamically, by token — which makes cost planning harder.
Zapier: the agents are a separate product at a separate price, drawing on their own "activity" quota — not on your task quota. The MCP integration does draw on it, though: two tasks per action.
Data handling and log retention
This part rarely comes up when choosing, yet for an auditable process it is what decides.
| n8n Cloud plan | Max. saved executions | Log retention |
|---|---|---|
| Starter | 2,500 | 7 days |
| Pro | 25,000 | 30 days |
| Enterprise | 50,000 | unlimited time |
The system deletes the log automatically after a while or when the storage limit is reached — whichever comes first — always starting with the oldest.
And an honest sentence from the documentation: at high disk saturation the provider takes a backup and restores the instance without execution data. They add that "because of the human steps involved, the alerting system is not perfect", and that if the warning arrives outside working hours, there may not be time to delete.
Then there is the memory limit: the entry plan gets 320 MiB, of which the software itself uses an average of 180. According to the documentation, with a complex workflow processing large amounts of data the instance may crash and become unreachable.
The lesson: if you need an auditable trail — GDPR, an accounting process — the entry plan's seven-day log is not enough. It has to be written out to external storage.
Version check: what has changed
Three things that make a 2024 comparison look immediately out of date:
- Make today counts in credits, not in "operations". According to the help pages, the credit replaced the operation as the name, with prices unchanged. If an article says operations, it is old.
- The platform called Integromat no longer exists under that name — it became Make.
- n8n Gateway credits and the Registered Community tier are both new, and most earlier comparisons do not mention them.
All five platforms are alive and running under their own names, with live price lists.
Common misconceptions
- "n8n is open source, so it is free." It runs under a fair-code licence, which is commercially restricted. You may use it free inside your own company, but you may not pass it on as a service.
- "Self-hosted I get everything." SSO, projects, sharing and Git version control ask for a paid licence key — self-hosted as well.
- "The cheapest plan is the cheapest." The billing unit decides. At twenty steps the "more expensive" plan can be twenty times better value.
- "The model cost is included." By default it is not. Either you pay it with your own key, or you buy a separate balance that expires.
- "I will look at the log later." On the entry plan it is seven days and then it is deleted — and in an emergency the provider may delete too.
- "The trigger consumes as well." Not on Zapier. A replay does, though, all over again.
What to take away
Three questions, in this order:
One: how many steps does one of your runs have, and how often does it run? This decides which billing model favours you — and the difference can be twentyfold.
Two: do you work alone or in a team? Alone, the self-hosted free version is enough. In a team, sharing, permissions and version control become paid, wherever you host.
Three: do you need an auditable trail? If so, the entry plans' log retention is too short, and you have to plan for external storage — which is cheaper to decide at the start than afterwards.
What to start with in the first place, and why the platform fee is the item that is orders of magnitude larger than the model cost, is worked out with numbers in the your first automation post; the pitfalls of automations running on a schedule in the background are collected in the scheduled agents post, one by one.