Zapier vs Make vs n8n: AI Automation Platforms Compared (2026)

Key takeaways

• Zapier bills a task for every successful action step, so a workflow with eight actions costs eight tasks per run; triggers, filters, Paths, Formatter and Delay steps cost nothing.
• n8n bills one execution per whole workflow run regardless of node count, which is why step-heavy pipelines are dramatically cheaper on n8n than on Zapier or Make.
• Make renamed operations to credits on 27 August 2025; standard modules stay at one credit per operation, and action modules bill one credit per input bundle.
• Since 15 June 2026, an AI by Zapier step costs 1x, 3x or 5x tasks by model tier, with new steps defaulting to Advanced (3x) and each tool call billed at the same multiplier.
• n8n Cloud stores data in Frankfurt and self-hosting puts the instance wherever you want — the strongest position of the three for GDPR-sensitive workloads.
• Zapier's Looping step caps at 500 iterations with one looping step per Zap, the clearest signal a workflow has outgrown the platform.

Zapier meters successful action steps, Make meters module calls, n8n meters whole workflow runs. Headline prices tell you nothing until you translate them into your own workflow shape.

On Zapier, a task is any successful action that runs. The help documentation is specific about what is free: all trigger steps, Filter and Paths steps, failed or halted actions, and utility apps including Formatter, Delay, Looping, Digest, Storage, Tables and Forms. Everything else counts, including actions inside error handler branches.

On Make the unit is a credit, renamed from "operation" on 27 August 2025. One operation equals one credit for standard apps, but action modules bill one credit per input bundle — Make's own example is deleting ten rows, which costs ten credits, not one. Triggers and search modules stay at one credit regardless of how much data they return.

On n8n an execution is "a single run of your entire workflow", and its pricing page states that step count and data volume do not matter. A twelve-node workflow costs the same as a two-node one. That single decision is why the cost curves diverge at scale.

The crossover point is a function of steps per run, not runs per month. The more work a single workflow does, the more n8n's flat execution pricing wins.

Per Zapier's pricing page (checked August 2026), Professional starts at $19.99 per month billed annually for 750 tasks, rising through $49 for 2,000 and $129 for 10,000. Team starts at $69 for 2,000 tasks and adds SAML SSO. Free accounts get 100 tasks and two-step Zaps only.

Per Make's pricing page (checked August 2026), Free gives 1,000 credits, two active scenarios and a 15-minute minimum interval. Core starts at $9 per month, Pro at $16 and Teams at $29, all beginning at 10,000 credits with unlimited scenarios.

Per n8n's pricing page (checked August 2026), Cloud Starter is EUR 20 per month billed annually for 2,500 executions, Pro EUR 50 for 10,000, and Business EUR 667 for 40,000 plus self-hosting with a licence key.

A workflow with one trigger and seven actions, run 1,000 times a month, consumes 7,000 Zapier tasks, roughly 7,000 Make credits, and exactly 1,000 n8n executions — a $129 Zapier tier against a $9 Make tier against Starter-tier n8n, before anyone adds an AI step.

All figures below come from vendor pricing, help and documentation pages checked in August 2026. Treat the billing unit, not the number, as the durable difference.

Only n8n offers a genuine self-hosting path; Make and Zapier are hosted services with regional variation at best. For EU teams processing personal data, that is often the whole decision.

n8n's Community Edition ships under the Sustainable Use License, a fair-code licence that is explicitly not OSI-certified open source. Internal business use costs nothing; hosting n8n for paying third parties requires a commercial licence. Run it in a Docker container on a German VPS and n8n has no access to that instance, so your data processing agreement is with the host. n8n Cloud stores hosted-plan data in Frankfurt.

Make operates regional zones and an organisation is bound to the one it was created in. The Make API documentation lists eu1.make.com, eu2.make.com, us1.make.com and us2.make.com as zone base URLs reflecting the infrastructure region; migrating between zones is not self-service. Make's security page claims SOC 2 Type II and SOC 3 audits and ISO 27001 certification for Enterprise.

Zapier's legal pages state compliance with GDPR, UK GDPR and CCPA, but do not document an EU-only storage region. If you face a residency requirement rather than a general GDPR obligation, get it confirmed in writing first — the gap between "GDPR compliant" and "stored only in the EU" is where procurement reviews stall. Our notes on EU AI Act Article 50 transparency rules cover obligations that apply wherever the automation runs.

All three ship first-party AI steps and all three price them above ordinary steps. Zapier's model is the most aggressively metered.

Since 15 June 2026, AI by Zapier steps are billed by model tier: Standard 1x tasks, Advanced 3x, Premium 5x, bring-your-own-key 1x. New steps default to Advanced, and tool calls carry the same multiplier, following (1 x model rate) + (tool calls x model rate). Zapier's worked example: a Premium step making two tool calls consumes 15 tasks in one run.

Make's AI Toolkit and AI Agents apps cost one credit per operation plus token-based credits on Make's built-in provider. Connect your own provider on a paid plan and Make charges only the operation credit while your model vendor bills tokens.

n8n's AI layer is node-based and LangChain-derived: an AI Agent node orchestrates, with sub-nodes for model, memory and tools. Self-hosted n8n adds no platform surcharge, and it unlocks the option neither competitor offers — running open-weight models locally through Ollama so prompt content never leaves your network. That pairs with best open-source AI models in 2026.

All three now speak Model Context Protocol, the emerging standard for how an AI agent reaches external tools. The differences are direction and billing.

Zapier MCP exposes Zapier's action library to any MCP-compatible client. Per Zapier's MCP documentation, it runs on all plans including Free and consumes two tasks per successful tool call from the same allowance your Zaps use. Failed calls consume nothing, and batch actions count individually: emailing three people is six tasks.

Make works both ways: its MCP server lets an external AI client run your scenarios, and its MCP client lets a scenario or Make AI Agent call tools hosted elsewhere. n8n does the same through dedicated nodes — the MCP Server Trigger exposes a URL that clients connect to, publishing only the tools you attach, while the MCP Client Tool node consumes external servers. New to the protocol? Start with our Model Context Protocol guide, then AI agents that take actions.

This is where the platforms differ most in daily operational cost. Make has the most explicit failure model, Zapier the most automatic, n8n the most configurable.

Zapier's Autoreplay, on Professional and above, replays steps that fail on transient errors up to five times, the final attempt landing roughly ten hours and thirty-five minutes after the first error. Filter and Paths steps are never replayed. Note the interaction: attaching a custom error handler turns Autoreplay off for that Zap, and runs with error handlers cannot be replayed manually. You get automatic retries or custom branching, not both.

Make attaches error handler directives to individual modules — Ignore, Resume, Commit, Rollback, Break and Retry. Break parks the failing bundle in the Incomplete Executions queue instead of killing the run, defaulting to three retry attempts fifteen minutes apart. Make also auto-retries incomplete executions caused by rate limit, connection and module timeout errors on an exponential backoff schedule.

n8n works at two levels. Nodes expose a Retry On Fail setting with Max Tries and Wait Between Tries in milliseconds — n8n's documentation suggests matching the wait to the target API, for example 1,000 ms for one request per second. At workflow level you designate an error workflow beginning with an Error Trigger node, which receives the failing workflow ID and error message.

Make and n8n treat data as bundles or items flowing through a graph. Zapier treats it as a linear run with optional branches, and that model has hard ceilings.

Zapier's Looping step is capped at 500 iterations, only one looping step is permitted per Zap, and nested loops are unsupported. Each action step after the loop consumes one task per iteration, so a loop running 500 times makes a single downstream action cost 500 tasks. Documented limits also include 100 steps per Zap and 1,000 fields per action step.

Make's flow control primitives are routers, filters, iterators, aggregators and repeaters, available even on Free. An iterator splits an array into bundles, an aggregator collects them back, and a repeater generates a fixed number of bundles from nothing, which is how most people paginate an API. Because action modules bill per input bundle, a fan-out of 200 items through three action modules costs 600 credits.

n8n's canvas is item-based, so most nodes run once per input item with no explicit loop construct, and item count does not change the price. That is why bulk data pipelines end up on n8n.

None of the three ships real version control on entry plans. n8n is the only one that connects to Git, and only on Business and Enterprise.

Zapier separates drafts from published Zaps: edit a draft without turning the live Zap off, then publish to replace it. Version rollback is available on Professional, Team and Company, and retention scales with tier — Zapier's pricing page lists roughly one month on Professional, six months on Team and one year on Enterprise, which also adds version comparisons and approval requests.

Make keeps scenario version history for up to 60 days. You restore from the version control in the editor, pick a version, then save manually — restoring does not save by itself. Scenario blueprints are exportable JSON you can check into a repository, the closest Make gets to source control.

n8n's source control and environments feature links an instance to a Git branch and uses a push-pull pattern to move work between environments. The Community Edition comparison excludes Git version control, environments, projects, custom variables, external secrets, log streaming and SSO from the free self-hosted edition; registering by email adds only folders, debug-in-editor and custom execution data. n8n 2.0, stable since December 2025, also split Save from Publish so saving an active workflow no longer immediately changes production.

Every platform has a code step, and its ceiling tells you how far the platform will carry you before you need real infrastructure.

Code by Zapier supports JavaScript on Node.js 22 and Python 3.10, with script timeouts of 10 seconds on Starter and 30 seconds on Pro, Team and Company, extendable to 10 minutes; billing for extended runtimes began 15 June 2026. The sandbox runs on AWS Lambda with a 6 MB input and output limit covering both code and data. Make's Code app charges two credits per second of execution time.

n8n's Code node supports JavaScript and Python via native task runners; Pyodide is a legacy feature that n8n 2.0 no longer supports. On n8n Cloud you cannot import external npm modules — only the crypto built-in and the moment package — and Python cannot import libraries at all. Self-hosted instances lift both restrictions through the NODE_FUNCTION_ALLOW_BUILTIN and NODE_FUNCTION_ALLOW_EXTERNAL environment variables. If your automation needs a specific library, that alone decides Cloud versus self-hosted.

Zapier wins on breadth, and that is the only reason many teams pay its prices. Its app directory advertised 9,000+ apps when checked in August 2026, Make's integrations page stated 3,000+, and n8n's listed 1,993, with the HTTP Request node promoted alongside named connectors for anything uncovered.

Raw counts overstate the gap: if your stack is Google Workspace, Slack, HubSpot, Stripe, Notion and a database, all three cover you. The count matters when you depend on a narrow vertical product — a practice management system, a regional payment provider, a niche CRM. That is where Zapier's directory is hard to replace and where n8n turns into hand-configured HTTP Request nodes. More in our AI productivity tools roundup.

Coverage also decides whether you need a general-purpose automator at all. If the destination systems already ship their own AI layer, that layer may be cheaper than routing through steps you pay for per run: our comparison of AI project management tools covers the built-in automation in Asana, monday and ClickUp, and AdCreative.ai vs HubSpot covers the same question for marketing and CRM workflows.

Governance features cluster at the top of each pricing table, and the gap between "the team can use it" and "IT will approve it" is usually two tiers.

Zapier puts SAML SSO on Team and above; Enterprise adds SCIM provisioning, domain capture, app and action restrictions and a Super Admin role, and Company and Enterprise accounts can set custom data retention between 7 and 30 days. Make gates teams, roles, SSO and 2FA enforcement to Teams and above. n8n gates SAML and LDAP SSO, projects, external secrets and log streaming to Business and Enterprise, and self-hosted paid instances run on licence keys that contact n8n's licence server daily — so an air-gapped deployment needs an explicit vendor conversation.

The signal to move is rarely capability. It is cost per run, review process and blast radius.

The first threshold is arithmetic: when the monthly platform bill exceeds a small VPS plus a developer's maintenance time, the no-code premium has stopped paying for itself. Self-hosted n8n is usually the intermediate step rather than a full rewrite, because it removes per-run billing without discarding the visual editor.

The second is structural: Zapier's 500-iteration loop ceiling, one-looping-step rule and 100-step limit are hard walls, and chaining four sub-Zaps to dodge a loop limit means the work is in the wrong medium. The third is process — once more than one person edits a workflow others depend on, you need branch-and-merge review, not a 60-day history and a manual save. The fourth is dependency: when an automation becomes something customers experience, you want tests, staged deploys and rollback the platform cannot fully provide.

Two counter-signals argue for staying put. If the value is that non-engineers can change it, code destroys that. And if it leans on twenty vendor integrations, replicating that connector surface is a commitment most teams underestimate.

The decision reduces to three questions: how many steps per run, who edits it, and where the data may live.

One pattern beats picking outright: keep Zapier for the handful of odd connectors nothing else supports, and run high-volume pipelines on self-hosted n8n where per-run billing does not apply. The integration surface and the execution engine are different problems.