DEV Community

dubleCC
dubleCC

Posted on • Originally published at heycc.cn

Build AI Workflows in 2026: n8n vs Make vs Zapier

#n8n #make #zapier #aiautomation

Originally published at heycc.cn. This is a mirrored copy — the canonical version is kept up to date at the source.

Build AI Workflows in 2026: n8n vs Make vs Zapier

If you want to wire an LLM into an automation — read an email, ask Claude or GPT to summarize it, post the result to Slack — you have three obvious no-code/low-code platforms to reach for. The marketing pages all say roughly the same thing, so the real decision isn't "which tool has AI." All three do. The decision comes down to two things almost nobody compares head-to-head: how each platform actually calls the model, and what billing unit it charges you per run.

Those two facts dominate everything else. The billing model decides whether a 15-step AI pipeline costs you single-digit dollars or hundreds a month. The integration model decides whether you can self-host the whole thing, store API keys in a vault, and drop into raw HTTP when the pretty node doesn't do what you need. Feature checklists rarely surface either, so most comparison posts leave you no closer to a decision. This one is organized around those two axes instead.

The comparison that actually matters

Each platform charges in a fundamentally different unit. That single fact — not the app count, not the UI — is what makes one of them cheap and another expensive for your specific workflow shape.

Tool Best for How it calls an LLM Billing unit Self-host?
n8n Developers wanting control, local LLMs, complex pipelines Native AI Agent node (LangChain + ReAct) with a separate Chat Model node, OR raw HTTP Request node, OR Ollama for local models Per full workflow execution: one run = one execution, regardless of node count Yes, free open-source Community Edition on GitHub
Make Visual builders, mid-complexity branching scenarios Dedicated Anthropic/OpenAI app module, OR generic HTTP module to the provider endpoint; key stored once in a Connection Per credit: most module actions cost roughly 1 credit, advanced/AI actions can cost more No (cloud SaaS only)
Zapier Non-technical users, widest app catalog, quick wins Built-in ChatGPT/Anthropic actions, "AI by Zapier," AI Actions, and Zapier MCP Per task: each successful action step = 1 task No (cloud SaaS only)

The billing column is the one to stare at. Make bills every step including filters and formatting modules. n8n bills the entire workflow as a single execution no matter how many nodes it contains. Zapier sits in between. It only charges for "work" action steps, so triggers and filters are free, but each AI call and each downstream write is its own task. For an AI pipeline that fans out across ten steps, this difference is enormous. We do the actual arithmetic below.

Decision flow comparing how n8n, Make, and Zapier call an LLM and bill per run

How you actually call an LLM in each

n8n

n8n gives you two clean paths. The high-level one is the AI Agent node, which is LangChain-powered and built on the ReAct (Reasoning + Acting) framework. The agent node doesn't contain a model itself. You attach a Chat Model node (an OpenAI Chat Model, Anthropic Chat Model, etc.) underneath it, plus optional memory and tool nodes. That separation is the whole point: swap the model without rebuilding the agent.

The low-level path is the plain HTTP Request node pointed at any provider's API. If a model ships a feature before n8n adds a node for it, you don't wait; you POST JSON directly.

The differentiator is local models. n8n publishes an official self-hosted AI starter kit, a Docker Compose template bundling n8n with Ollama (local LLMs), Qdrant (vector store), and PostgreSQL. You can run an entire RAG-style AI workflow with zero data leaving your machine — relevant if you handle anything regulated.

The canonical pattern from n8n's AI integration docs is the one you'll build first: a trigger (say Gmail) feeds data into an AI Agent / Chat Model node, and the model's output maps into a downstream action node: a Slack post or a Notion page. Trigger to model to destination.

Make

In Make you call an LLM one of two ways. The clean way is a dedicated app module (the Anthropic Claude module or the OpenAI module), where you pick the model, set the prompt, and Make handles the request shape. The flexible way is the generic HTTP module pointed straight at the endpoint, e.g. api.anthropic.com/v1/messages. Either way, the API key lives once inside a saved Connection and is reused across every module that needs it, so you're not pasting keys into individual steps.

One sharp edge worth knowing: the Anthropic Messages API returns an array of content blocks, not a single string, and a response can include one or more tool_use blocks ahead of the final text block, so you can't just map a fixed array index. The reliable Make-native fix, per a Make community thread, is to run the array through an Iterator and follow it with a Filter that only passes bundles where type is text. That detail trips up first-timers who expect a flat text field. This is the kind of thing the dedicated module hides for you but the raw HTTP module does not.

Zapier

Zapier leans on built-in actions for the ChatGPT (OpenAI) and Anthropic (Claude) apps, plus AI by Zapier and AI Actions for lighter inline tasks. The 2025-era addition worth your attention is Zapier MCP, which connects Claude or ChatGPT directly to Zapier's app catalog. Zapier's own MCP guide describes access to "thousands of apps" and 30,000+ actions, with built-in governance: authentication runs through Zapier with keys encrypted and rotated, admins set per-app and per-action toggles, and there's a centralized audit log. The model can take actions without credentials ever being exposed to it. If your goal is "let an agent touch lots of SaaS apps safely," Zapier MCP is a strong fit here, precisely because the app catalog is the widest of the three.

Cost: where the billing models diverge

Pricing below was checked against each vendor's official pricing page on 2026-06-28 (see the dated verification section and Sources). Read the table together with the per-tier notes, because the three columns are not directly comparable line-for-line, because each platform sells a different unit, and even the billing basis differs (n8n quotes annual-billed monthly rates, Make quotes monthly rates, Zapier quotes annual-billed monthly rates). Each column is labeled so the basis is traceable.

Tier n8n Cloud (annual-billed monthly rate) Make (monthly rate) Zapier (annual-billed monthly rate)
Free Self-hosted Community Edition is free & open-source $0 (1,000 credits/mo, 15-min min interval) $0 (100 tasks/mo)
Entry paid Starter €20/mo, 2,500 executions Core $12/mo, 10,000 credits Professional ~$19.99/mo, 750 tasks
Mid Pro €50/mo, 10,000 executions Pro $21/mo, 10,000 credits Team ~$69/mo, 2,000 tasks
Higher Business €667/mo, 40,000 executions (self-hosted) Teams $38/mo, 10,000 credits (custom above Team)
Top Enterprise (custom) Enterprise (custom) Enterprise (custom)

Read the units carefully, because they're not comparable line-for-line:

  • n8n counts one execution per whole workflow run. A 2-step workflow and a 50-step workflow each cost exactly one execution. Self-hosted Community Edition has no per-execution license fee at all — your only cost is the server it runs on.
  • Make counts credits. The published guidance is that most module actions cost about 1 credit, but advanced and AI-provider actions can cost more, and Make has been migrating its billing language from "operations" to "credits," so check your specific plan generation before assuming a flat 1-credit rate. Filters and formatting steps still consume credits. Every paid tier in the table above includes 10,000 credits/month at the listed price. The Free plan also caps you to a 15-minute minimum scheduling interval versus 1 minute on paid plans.
  • Zapier counts one task per successful action step. A 5-step Zap consumes 5 tasks per trigger run. Triggers and filters/searches that perform no action don't count, but every AI call and every write does.

A worked example (this is the whole thesis)

Take one identical AI pipeline and run it on all three: a trigger, one LLM call, two formatting/parsing steps, and two writes (Slack + a database) — 5 billable action steps per run — executed 1,000 times a month. Here is the actual arithmetic, not a hand-wave:

Platform What gets billed Per-run cost in units At 1,000 runs/mo Cheapest plan that covers it
n8n 1 execution per whole run 1 execution 1,000 executions Starter €20/mo (2,500 executions), well inside it; self-hosted CE = €0 in license
Make 1 credit per action (≈5 credits/run, more if AI action costs extra) ~5 credits ~5,000 credits Core $12/mo covers it (10,000 credits), but you've spent half your allotment on one workflow
Zapier 1 task per action step 5 tasks ~5,000 tasks Team ~$69/mo (2,000 tasks) is still short; you'd need a higher tier or a task add-on

The pattern is the point. On n8n, doubling the step count to a 10-step pipeline changes the bill by zero. It's still 1,000 executions. On Make and Zapier, every step you add multiplies straight through your run count: a 10-step version is ~10,000 credits on Make and ~10,000 tasks on Zapier at the same 1,000 runs. That is why "pennies vs hundreds" isn't rhetoric: for step-heavy AI pipelines run at volume, per-execution billing wins decisively, and self-hosted n8n removes the per-run license entirely. If you want to model this for your own provider token spend on top of platform fees, see our guides to controlling LLM API cost and comparing the Claude, GPT, and Gemini APIs.

The flip side: for a simple, low-volume automation (a 2-step Zap run a few hundred times a month), Zapier's free or Professional tier is the least friction and the math barely matters. The billing-unit advantage only compounds as steps and volume grow.

Which one fits your workflow shape

  • Choose n8n if you want self-hosting, local/offline models, raw-HTTP control, or you're building step-heavy pipelines at volume where per-execution billing crushes the per-step competitors.
  • Choose Make if you want a polished visual canvas for mid-complexity branching logic and your workflows are not so step-heavy that per-credit billing hurts.
  • Choose Zapier if breadth and speed for a non-technical user matter most, your workflows are short, or you specifically want an agent to safely take actions across a very wide app catalog via MCP.

Where the prices came from

Each figure is tied to a vendor pricing page read on 2026-06-28. The n8n Cloud tiers (Starter €20 / Pro €50 / Business €667, with the executions per tier) are the annual-billed monthly rates from n8n.io/pricing. The month-to-month rates run higher. The Make tiers (Core $12 / Pro $21 / Teams $38, each including 10,000 credits, plus the free tier's 1,000 credits and 15-minute interval) are from make.com/en/pricing. The Zapier rates come from zapier.com/pricing: Professional (~$19.99/mo, 750 tasks) and Team (~$69/mo, 2,000 tasks), both annual-billed. The one figure most likely to drift is Make's: its billing unit is mid-migration from "operations" to "credits," so a given plan generation may count differently than the table assumes.

Sources

Top comments (0)