n8n vs Zapier vs Make for AI workflows: what they really cost

In n8n vs Zapier vs Make, what an AI workflow costs depends more on the unit each tool bills than on the plan price. Zapier counts a task for every successful action and Make a credit for every module operation. n8n counts one execution per workflow run, however many steps it has, so adding a step raises the bill only on Zapier and Make.

For the eight-step AI triage workflow priced below, the cheapest plans covering 30,000 runs a month cost $769 a month on Zapier, $237 on Make and $120 on n8n Cloud, billed annually at US prices as of 22 October 2026. At 3,000 runs Make is cheapest, at $38 against $50 for n8n Cloud and $189 for Zapier. Model tokens add between $1 and $210 a month on top.

Below are the counting rules, the arithmetic, AI agents, n8n’s licence, hosting, testing and when to move into code. We haven’t run these tools; the figures come from vendor pages and help centres.

n8n vs Zapier vs Make: what each one counts

Zapier bills tasks. A task is a successful action step. Triggers, polling, Filter and Paths steps, Formatter, Delay, Looping and steps that error don’t use tasks, but steps on an error-handler path do.¹

Make bills credits, which replaced operations at unchanged prices. An operation is one module run on one bundle of data, and most cost 1 credit.² A trigger uses one operation per check. Later modules run once per bundle, so a search that returns three records runs the rest of the scenario three times. Routers and error-handler modules are free.³

n8n bills executions. An execution is one production run of a whole workflow, whatever its number of steps. Manual test runs, sub-workflow runs, error-workflow runs and polls that find nothing don’t count.⁴ Each turn of a chat agent counts as one execution.⁵

Zapier pricing, Make pricing and n8n pricing as of 22 October 2026

Prices are monthly, in US dollars, as each vendor showed them for the United States on 22 October 2026. Make and n8n price by country. Make shows visitors from Germany the Core plan at $9 for 10,000 credits billed annually, against $12 for the United States, and shows visitors from India a single Make Plan instead of Core, Pro and Teams.³ n8n’s Starter and Pro 10K figures are the same in euros, pounds and dollars, but Business costs $960 a month in the United States against €800.⁵ Zapier converts its dollar prices when you pay in another currency.⁶

The limits come from the vendors’ billing articles.⁷⁸ Enterprise plans are priced by quote.

Among other Zapier alternatives, Activepieces also bills one credit per run, plus 2 to 20 credits for an AI step on its own models.⁹

Worked example: an AI triage workflow at 3,000 and 30,000 runs

The workflow triages support tickets in eight steps. A webhook receives the ticket and one LLM call classifies and summarises it. Two lookups fetch the customer from the CRM and their plan from billing, and a branch separates urgent tickets. Two writes update the ticket and log the result, and a Slack message goes to the right channel.

We assumed every run passes all eight steps, each lookup returns one record, nothing fails, and the LLM call uses your own API key in an ordinary app step. By each vendor’s rules, a run is then 6 tasks on Zapier, where the trigger and Paths step are free, 7 credits on Make, where only the router is free, and 1 execution on n8n. Plans are the cheapest tier covering the volume, at US prices billed annually, with monthly billing in brackets.

Make’s 21,000 credits sit just over its 20,000 tier. Staying on Core 20K at $21 and buying 1,000 extra credits by hand adds 1,000 × ($21 ÷ 20,000) × 1.25 = $1.31 a month.⁸ In n8n vs Make, the cheaper option flips more than once between these volumes. At 8,000 runs, n8n Cloud Pro 10K ($50) beats the Core 80K plan ($72) that Make’s 56,000 credits need.

n8n Starter’s 2,500 executions don’t cover 3,000 runs, and n8n doesn’t say what happens past a Starter or Pro quota. Self-hosted Community has no quota or licence fee but needs a server and upkeep, which we haven’t priced.

On AI by Zapier’s Premium tier, the default for new steps on paid plans, the LLM step costs 5 tasks and the run 10. Zapier then pays for the tokens, and the plans rise to Professional 50K ($289) at 3,000 runs and 300K ($1,069) at 30,000.¹⁰

Model tokens are billed separately. We assumed 2,000 input and 300 output tokens per run and no reasoning tokens, at list prices from our comparison of LLM API prices as of 8 October 2026.

At 30,000 runs on a small model, the platform fee is 11 times the token bill on n8n Cloud and 73 times on Zapier. On a mid-tier model, n8n Cloud’s fee falls below it. Reasoning tokens bill as output and would raise both rows.

Where model tokens are billed: your own key or the platform’s credits

All three accept your own API key, and each also resells model access.

Zapier. AI by Zapier costs 1 task per run on Standard models such as GPT-5 mini, 3 on Advanced and 5 on Premium, and each tool call adds the same multiple again. Your own AI account counts like Standard, and a run pauses for approval at 75 tasks, adjustable to 500.¹⁰ Enterprise plans can route supported AI processing through the customer’s own AWS Bedrock account.⁶

Make. Make’s AI Provider, available on every plan in AI Agents and the AI Toolkit, adds token credits to the operation. Since 25 August 2026 its Large tier (GPT-5 mini) charges a credit per 3,616 input or 452 output tokens.¹¹ For our example that is 2,000 ÷ 3,616 + 300 ÷ 452 = 1.22 extra credits a run, which fits the same Core tiers. With your own provider on a paid plan, the step stays at 1 credit.²

n8n. Gateway credits on Cloud Starter and Pro run OpenAI, Anthropic, Google and other models from a prepaid balance. n8n’s service pricing page matches the 8 October list prices for the models we checked, such as $0.10 and $0.50 per million tokens for GPT-6 Luna. Top-ups expire after 12 months, and Gateway credits aren’t offered on Cloud Enterprise or self-hosted n8n.¹²

Zapier Agents, Make AI Agents and n8n’s AI Agent node, with MCP

Zapier. Zapier Agents bills activities, a quota separate from tasks, with 400 a month free or 1,500 for $33.33 a month billed annually.¹³ Zapier is moving Agents into AI by Zapier steps, which bill tasks, and hasn’t set a date to turn Agents off.¹⁴ Zapier MCP lets clients such as Claude and Cursor call over 40,000 actions on any plan, at 2 tasks per successful call. The beta MCP Client app calls remote MCP servers from a Zap; its docs don’t say how that is billed.¹⁵

Make. Make AI Agent (New), released on 2 February 2026 and in open beta, runs agents inside scenarios on every plan. A run costs 1 credit per operation plus token credits on Make’s provider, and each tool call adds operations. Tools can be modules, scenarios, MCP server tools or other agents.¹⁶ Make MCP server exposes on-demand scenarios to MCP clients, with run scopes on every plan and management scopes on paid plans. Its MCP Client app, in open beta, calls remote servers from a scenario, at a price Make says may change.¹⁷

n8n. The AI Agent node pairs a chat model with tool sub-nodes. The MCP Client Tool node hands an agent an external server’s tools, and the MCP Client node uses them as ordinary steps. The MCP Server Trigger node exposes a workflow’s tools as an MCP server, and instance-level MCP lets clients search, run and build the workflows an admin exposes.¹⁸ Tool calls stay inside one execution, but each chat turn is a new one, and calls to an MCP Server Trigger’s production URL are production executions.

n8n self-hosted and the Sustainable Use License

n8n’s Community edition is free to self-host. SSO, Git and environments need a Business or Enterprise licence key, and the pricing page puts external secrets, log streaming and audit logs on Enterprise.⁵ The code is under the Sustainable Use License, which n8n’s docs now call its Community license, except files with “.ee.” in their names, which need an n8n Enterprise License.¹⁹²⁰

The licence lets you use or modify n8n “only for your own internal business purposes or for non-commercial or personal use”, and distribute it only free of charge for non-commercial purposes.¹⁹ n8n’s licence FAQ turns that into one test, which is who controls the workflow logic.²¹ As of 22 October 2026 it allows building and running workflows for clients on your own instance, provided they can’t create or edit them. It also allows n8n behind the scenes in your own product, where users trigger workflows you built and can connect their own accounts. It rules out hosting n8n as a service, white-labelling it, forking it into your own automation product, and using enterprise features without a licence key. End users also mustn’t build or configure workflows through your product, whether by a custom UI, the API, MCP or an AI agent acting for them.

Under that FAQ, the triage workflow in this post can run behind your product on self-hosted n8n. A feature that lets your customers build their own automations needs a commercial licence even if they never see n8n’s editor. Putting the editor itself in front of them needs n8n’s separate OEM offering, with n8n’s branding visible.²²

n8n’s other pages don’t all agree with the FAQ. Its OEM page says back-end use “runs on a regular Enterprise license”, and its OEM docs say back-end use is available on all paid plans with no separate agreement.²² This isn’t legal advice. If a feature depends on self-hosted n8n, get n8n’s answer in writing and have a lawyer check the edge cases.

Hosting and data residency

Check where the model endpoint runs as well as the platform. Zapier stores customer data in AWS US-East-1 and offers region-specific residency only by separate agreement.²³ Make runs on AWS in the EU and North America, with each organisation in the US1, US2, EU1 or EU2 data zone. Make warns that an EU zone doesn’t keep every AI call in the EU. OpenAI models on its AI Provider run in the US, your own OpenAI, Anthropic or Gemini connection follows that provider’s routing, and Microsoft Foundry or other OpenAI-compatible endpoints keep EU processing in the EU.²⁴ n8n Cloud stores data in Frankfurt, and self-hosted n8n stays wherever you run it.⁵

Version control, testing and environments

The table draws on each vendor’s version, replay and source-control docs.²⁵²⁶²⁷ Zapier says full-Zap tests of migrated agents currently use tasks like production runs.¹⁴ On n8n’s Community edition you can export workflows as JSON and commit them yourself.²⁸

Error handling and retries

Zapier. Steps on an error-handler path use tasks. Autoreplay, on Professional and above, retries a failed step up to five times over about 10 hours 35 minutes. Replays from Zap history rerun only the failed step, so a model call that already succeeded isn’t billed again, while a full replay from the trigger bills every successful step twice.²⁹

Make. Error-handler modules are free. With incomplete executions turned on, which they aren’t by default, Make retries rate-limit, connection and timeout errors eight times over 7 hours 51 minutes, starting each retry from the module that failed.³⁰

n8n. Each node has Retry On Fail and an On Error setting that stops, continues or routes the error to its own output. Error workflows, which start with an Error Trigger, don’t count towards the quota.³¹⁴

When to build a custom AI agent or move a workflow into code

As a rule of thumb, keep a workflow in the tool until two of these signals appear.

1. Volume and cost. The yearly fee passes what a small service would cost to build and run. At 30,000 runs the example costs $9,228 a year on Zapier, $2,844 on Make and $1,440 on n8n Cloud.
2. Branching depth. More than two levels of nested branches, or loops inside branches, which also multiply tasks and credits.
3. Latency. A user waits on the result. Zapier Professional polls every 2 minutes, Make’s paid plans run schedules at most once a minute, and n8n Cloud queues runs beyond 5 concurrent executions on Starter and 20 or 50 on Pro.⁶³³²
4. Tests and code review. Changes need unit tests, evals or a reviewer before they ship. Of the three, only n8n lists Git version control, on Business and Enterprise.
5. Secrets and compliance. Credentials must live in your own vault, data must stay in one region, or auditors want logs your plan lacks. Zapier’s audit log starts on Team, while Make’s and n8n’s are Enterprise features.
6. Lock-in. The workflow has become part of your product. Each tool has its own format, so leaving means a rebuild, and hosted builders get retired, as OpenAI’s Agent Builder will be on 30 November 2026.

A background workflow owned outside engineering and costing a few hundred dollars a year should stay put. We say the same about AI itself: if a form or a database query solves the problem, use that. If a Zap does the job, keep the Zap. When the signals stack up, moving a workflow into a tested service with evals is work our AI engineering team takes on.

Frequently asked questions