Zapier vs Make vs n8n vs Pipedream: Best AI Workflow Platform (2026)

Updated: August 4, 2026

For most 2026 AI workflow projects, Zapier wins on breadth of integrations, Make wins on visual complexity per dollar, n8n wins if you self-host and want deep AI-agent primitives, and Pipedream wins for developer-heavy pipelines that need to run arbitrary Node or Python. The right pick depends on how you charge. Zapier bills per task, Make per credit, n8n per execution, and Pipedream per compute-second. Honestly, that single billing detail tends to decide the winner faster than any feature comparison, and it's the first thing I check before I benchmark anything else.

• Zapier charges per successful action step (task), Make charges per credit since the August 2025 rebrand from "operations", n8n charges per workflow execution, and Pipedream charges per compute-second. The billing unit alone changes total cost by 5–20x on the same workflow.
• All four platforms shipped native AI/LLM nodes in 2025 with OpenAI, Anthropic Claude, and Google Gemini support. Zapier's MCP server and n8n's Tools Agent node are the two most agent-ready primitives right now.
• n8n is the only self-hostable option, released under the Sustainable Use License. It's free for internal use with unlimited executions, and a commercial license is required if you resell it as a hosted service.
• n8n raised a $180M Series C in October 2025 at a $2.5B valuation. Pipedream was acquired by Workday in November 2025, with its self-serve tier preserved through at least June 2026 but longer-term uncertainty for external customers.
• For teams already fluent in a general-purpose language, Pipedream's code-first workflows and n8n's Function node collapse the "let me exit to a Lambda" step that Zapier and Make still require.
• Real cost isn't the sticker price. It's the platform's billing granularity multiplied by your workflow's step count. A 10-step Zap costs 10 tasks per run; a 10-step n8n workflow costs 1 execution.

The four platforms at a glance

Before the deep dive, here's the shape of the field as of August 2026. I've listed only the dimensions I actually reach for when I'm scoping a project: sticker price, billing unit, AI-native features, and the license terms that decide whether I can host the thing on my own infrastructure.

Dimension	Zapier	Make	n8n	Pipedream
Billing unit	Per task (action step)	Per credit (since Aug 2025)	Per execution (whole workflow)	Per compute-second at 256MB
Free tier	100 tasks/mo, 2-step Zaps	1,000 credits/mo, 2 scenarios	Community Edition, unlimited (self-host)	100 credits/mo, 3 workflows
Entry paid tier	$19.99/mo (750 tasks)	~$10.59/mo (10,000 credits)	$24/mo Cloud (2,500 executions)	~$29/mo (2,000 credits)
Self-hostable	No	No	Yes (Sustainable Use License)	No
Native LLM providers	OpenAI, Claude, Gemini via AI Step	OpenAI, Claude, Gemini canvas nodes	OpenAI, Claude, Gemini, Mistral, Groq, DeepSeek, xAI, Ollama, and more	Any provider via code + Connect
Agent primitive	Zapier Agents + MCP server	Make AI Agents (beta)	Tools Agent node (LangChain JS)	Managed MCP servers
Best for	Non-technical teams, breadth	Visual multi-branch flows	Self-hosted agents, complex data	Code-first serverless workflows

The rest of this article unpacks each row, because the numbers above hide the interesting decisions. A "task" in Zapier and a "credit" in Make don't count the same things, and Pipedream's compute-second model punishes long-running LLM calls in ways the others just don't.

How each pricing model actually works in 2026

Every one of these platforms will happily sell you a plan denominated in a unit you don't fully understand, so this section pays for itself. The core insight: your true cost is the price per unit multiplied by units consumed per run, multiplied by runs per month. Getting the middle term wrong is where most teams overspend.

Zapier: per successful action step

Zapier's task is a single successful action step. A 10-step Zap that runs 100 times consumes 1,000 tasks. The trigger is free, filters are free, but every action that executes counts. Failed steps don't charge. According to Zapier's current pricing page, the Professional tier is $19.99/mo billed annually for 750 tasks with unlimited Zaps and premium apps. AI Step calls and MCP tool invocations draw from the same task pool, which (I've noticed) surprises more than one team I've worked with.

Make: per credit since the August 2025 rebrand

On August 27, 2025, Make replaced "operations" with "credits" at a clean 1:1 mapping, then quietly repriced overages roughly 25% higher two months later. A credit still counts one module execution, so the "each step costs money" model matches Zapier's, not n8n's. The Core tier is around $10.59/mo for 10,000 credits, which is why Make consistently looks cheaper than Zapier for high-volume flows.

n8n: per execution, not per step

This is the model that changes the arithmetic. In n8n Cloud, one execution equals one full workflow run, no matter how many nodes fire inside it. A 40-node workflow that runs 1,000 times a month is 1,000 executions, and the $24/mo Starter tier's 2,500 executions cover 2.5x that. Self-hosted Community Edition has no execution cap at all. You pay for the VM, not the runs.

Pipedream: per compute-second at 256MB

Pipedream bills the compute time your workflow actually consumes, at 256MB base memory. This is beautiful for short webhook-to-Slack pipes and painful for anything that waits on a slow LLM. A 20-second Claude call at 512MB burns four credits every run. Free tier is 100 credits/month plus 1M AI tokens; Basic runs about $29/mo for 2,000 credits.

AI and LLM support compared

All four platforms shipped serious AI features in 2025 and iterated hard through 2026. The differences aren't whether they support Claude or GPT-4o (they all do), but how much orchestration lives inside the platform versus how much you push out to code.

Zapier: MCP server plus Zapier Agents

Zapier's 2026 pitch is that its catalog of 30,000+ actions is now exposed to any external LLM via its Zapier MCP server. Point Claude Desktop or a custom agent at the MCP endpoint and it can execute Salesforce updates, Google Calendar events, or Notion writes without you writing a single integration. Inside Zapier itself, "Zapier Agents" (the product formerly called Central) lets you describe an autonomous flow in natural language and Copilot builds the Zap.

Make: Maia and AI Agents beta

Make's "Maia" is a conversational scenario builder that generated most of the demo-video screenshots you've seen in the past year. Its AI Agents feature is still marked beta, which in practice means production teams I know use Make for the canvas and delegate the agent loop to something else.

n8n: the Tools Agent node

n8n 2.0 shipped in January 2026 and doubled down on being the LangChain of no-code. The Tools Agent node wraps LangChain JS's tool-calling loop natively. You connect an LLM node (OpenAI, Anthropic, Gemini, Mistral, Groq, DeepSeek, xAI Grok, Moonshot Kimi, Ollama for local, OpenRouter for gateway, or Vercel AI Gateway), attach any n8n node as a tool, and the agent orchestrates. I've covered the mechanics (and the failure modes) in my earlier piece on why n8n AI Agents don't call their tools and the follow-up on infinite loop causes when tools overlap. If you want to wire n8n's MCP surface to Claude Code, my n8n MCP tutorial walks the setup.

Pipedream: Connect and managed MCP

Pipedream's differentiator is that every workflow step can be arbitrary Node or Python, so "add Claude" is just npm install @anthropic-ai/sdk. Pipedream Connect embeds Pipedream's integration catalog inside your own product, and its managed MCP servers expose 10,000+ tools across 3,000+ APIs, a direct competitor to Zapier's MCP surface.

Is n8n really free? License, self-hosting, and the fair-code trade-off

Yes, n8n's Community Edition is genuinely free to self-host with unlimited workflows and executions. But it's released under the Sustainable Use License, which is fair-code, not OSI-certified open source. The distinction matters in three cases. First, if you plan to sell hosted n8n as a service to third parties, you need a commercial license. Second, SSO, LDAP, audit logs, external secrets, and workflow version control require an Enterprise license key even when you self-host. Third, some engineering leadership shops that enforce OSI-only licensing (some finance and defense teams, notably) can't ship n8n on that basis alone.

For everyone else (internal automations, agency work billed as consulting rather than hosting, personal projects), the Community Edition is exactly as free as it looks. A $10/month Hetzner VPS runs it fine for small teams. A 4-core node with Postgres handles thousands of executions per hour once you tune the queue-mode worker settings.

Is Make better than Zapier? Where each one actually wins

Both platforms are SaaS-only, both bill per step, and both have serious 2026 AI stories. The differences that matter in practice are the canvas, the app catalog, and the error-handling model.

Make wins on visual complexity. Its canvas handles multi-branch, parallel, and iterator patterns natively. Zapier's linear step model punts most of that to Paths (a paid feature) and gets clumsy above five branches. If your workflow diagram would need arrows crossing each other on a whiteboard, Make renders it cleanly and Zapier doesn't.

Zapier wins on catalog breadth. 7,000+ integrations to Make's roughly 2,000. In the long tail (niche CRMs, regional payment processors, industry-specific SaaS) Zapier is more likely to have the app you need without you writing a custom HTTP request block.

Make wins on price per step. A workflow that consumes 10 credits per run costs Make users roughly 10% of what the same 10 tasks costs Zapier users on comparable tiers. This gap widens at scale.

Zapier wins on non-technical usability. The linear Zap format, natural-language Copilot, and template gallery mean marketing teams and ops staff can ship without training. Make's visual model requires a mental shift that trips up first-time users.

Both are equivalent on AI. ChatGPT, Claude, and Gemini nodes exist on both, both shipped agent primitives in 2025-2026, and both are catching up to n8n's LangChain-native depth.

Which is cheaper: n8n, Make, or Zapier?

For any workload above the free tiers, self-hosted n8n is cheapest by a wide margin, followed by Make Cloud, with Zapier meaningfully more expensive per equivalent unit of work. But the ranking flips at low volume, because self-hosting has a fixed operational cost floor.

Here's a concrete model. Assume a workflow that runs 5,000 times a month and executes 8 action steps per run.

• Zapier: 5,000 × 8 = 40,000 tasks. That's the Team tier at $103.50/mo billed annually, or higher.
• Make: 5,000 × 8 = 40,000 credits. Pro tier at roughly $18.82/mo covers it with credits to spare.
• n8n Cloud: 5,000 executions on the Pro tier at $60/mo (10,000 executions included).
• n8n self-hosted: $10/mo VPS. No execution cap.
• Pipedream: If each run takes 6 seconds at 256MB, that's roughly 1,000 credits. The Basic tier at $29/mo covers double.

The n8n self-hosted number ignores operational overhead (backups, upgrades, monitoring) and it's real overhead. Budget an hour a week if you're doing it seriously, which at a $150/hr engineering rate is $600/mo of hidden cost. That inversion is why Make usually wins the "cheap and hands-off" bracket, and self-hosted n8n wins only once your volume covers the operational tax.

What replaces Pipedream after the Workday acquisition?

On November 19, 2025, Workday announced the acquisition of Pipedream. The deal closed in January 2026. Workday committed to keeping the self-serve tier available through at least June 2026, but long-term posture for non-Workday customers hasn't been clarified as of this writing.

If you're on Pipedream today and worried about the exit ramp, the closest functional replacements are:

• n8n's Function and Code nodes if you want to stay code-first inside a visual builder.
• Inngest, Trigger.dev, or Hatchet if you were mostly using Pipedream as a serverless queue with retries. I compared those three in the background job queue guide.
• Cloudflare Workflows or Temporal if you want durable execution and don't mind self-managing more. My deeper piece on LangGraph and Temporal for durable agent pipelines covers the trade-offs.

I don't recommend a panic migration. Workday isn't obviously going to kill self-serve. The acquisition rationale looked more like MCP-and-agents than lock-in. But if you're building a mission-critical integration surface, having a documented exit plan is now table stakes.

How I pick between them: a decision framework

The way I actually make this call on a client project, ignoring hype:

1. Does someone non-technical need to edit the workflow? If yes, Zapier. Everything else has a learning curve that costs more than the license premium.
2. Do you have hard data-residency, air-gap, or OSI-only license requirements? Only n8n self-hosted qualifies. Skip the rest.
3. Is the workflow multi-branch, parallel, or iterator-heavy? Make renders that model natively without extra cost. Zapier makes you pay for Paths and still gets awkward past five branches.
4. Are you building an AI agent with dozens of tools? n8n's Tools Agent node has the most mature primitive. Zapier's MCP surface is a strong second if you want the agent runtime elsewhere.
5. Is the workflow mostly custom code with a webhook trigger? Pipedream if you're willing to accept the Workday uncertainty; n8n's Code node if you're not.
6. Is monthly volume above 10,000 runs? Model self-hosted n8n TCO carefully. It usually wins, but only if you have someone who'll actually maintain it.

The mistake I see most often is teams optimizing for the wrong dimension. A three-person marketing team that picks n8n self-hosted because "it's free" spends more on the DevOps contractor than they would have on Zapier for years. A 50-person growth team paying Zapier by the seat for 20 users would save five figures a year moving to Make.

Same lead-enrichment workflow on all four platforms

To make the differences concrete, here's the same workflow (new HubSpot contact → enrich with Clearbit → summarize with Claude → post to Slack) sketched in each platform's idiom.

Zapier: four steps in a linear Zap. Trigger: HubSpot New Contact. Action 1: Clearbit Person Enrichment. Action 2: AI Step (Claude 3.5 Sonnet) with a prompt template referencing enrichment output. Action 3: Slack Send Channel Message. Consumes 3 tasks per run; the trigger is free.

Make: same four modules on the canvas, with an optional filter between Clearbit and Claude to skip low-confidence enrichments. Uses 3 credits per run. The canvas visualization makes the branching option trivial to add later.

n8n: HubSpot Trigger → HTTP Request (Clearbit) → LangChain LLM node (Claude) → Slack. One execution per contact. If you want it to be an agent ("look up the company on LinkedIn if Clearbit misses"), swap the linear chain for a Tools Agent node with three tools.

Pipedream: HTTP webhook trigger (HubSpot posts to Pipedream), followed by a Node code step that calls Clearbit's REST API, an @anthropic-ai/sdk step, and a Slack step. Roughly 4-8 seconds of compute per run at 256MB.

// Pipedream Node step calling Claude
import Anthropic from "@anthropic-ai/sdk"

export default defineComponent({
  props: {
    anthropic: { type: "app", app: "anthropic" },
  },
  async run({ steps }) {
    const client = new Anthropic({ apiKey: this.anthropic.$auth.api_key })
    const enrichment = steps.clearbit.$return_value
    const msg = await client.messages.create({
      model: "claude-3-5-sonnet-latest",
      max_tokens: 400,
      messages: [{
        role: "user",
        content: `Summarize this lead in 2 sentences for a sales rep:\n\n${JSON.stringify(enrichment)}`,
      }],
    })
    return msg.content[0].text
  },
})

The same logic in an n8n Function node:

// n8n Function node - assumes Anthropic credentials configured on the workflow
const enrichment = $input.first().json
const response = await this.helpers.httpRequestWithAuthentication.call(
  this,
  "anthropicApi",
  {
    method: "POST",
    url: "https://api.anthropic.com/v1/messages",
    body: {
      model: "claude-3-5-sonnet-latest",
      max_tokens: 400,
      messages: [{
        role: "user",
        content: `Summarize this lead in 2 sentences for a sales rep:\n\n${JSON.stringify(enrichment)}`,
      }],
    },
    json: true,
  },
)
return [{ json: { summary: response.content[0].text } }]

Notice how similar the code paths are. The platform difference is entirely in what happens around the code: trigger routing, credential storage, retries, observability, billing. That's the actual product each company is selling.

Frequently Asked Questions

Is n8n really free?

Yes, the n8n Community Edition is free to self-host under the Sustainable Use License with unlimited workflows and executions. You need a commercial license only if you resell n8n as a hosted service, and Enterprise features like SSO and audit logs require a paid license key.

Is Make better than Zapier in 2026?

Make is better for visually complex, multi-branch scenarios and for anyone cost-sensitive at scale. It's roughly 5-10x cheaper per step. Zapier is better for non-technical users, for the widest app catalog, and for MCP-based agent orchestration.

Which is cheaper: n8n, Make, or Zapier?

Self-hosted n8n Community Edition is cheapest above roughly 5,000 executions per month, once operational overhead is priced in. Make Cloud is cheapest for hands-off SaaS use across most tiers. Zapier is the most expensive per equivalent unit of work but often justifies the premium on non-technical usability.

What is the best Zapier alternative in 2026?

Make for visual multi-branch flows on SaaS, n8n for self-hosted AI-agent workflows, and Pipedream for code-first serverless pipelines, though Pipedream's future is uncertain post-Workday acquisition. Match the choice to your technical depth and your billing tolerance.

What replaces Pipedream after the Workday acquisition?

Workday committed to preserving the self-serve tier through at least June 2026, so no immediate migration is required. If you're planning ahead, n8n's Function and Code nodes cover most Pipedream use cases inside a visual builder, and Inngest or Trigger.dev cover code-first serverless queue patterns.

Can n8n replace Zapier for a small team?

Yes for technical teams comfortable with a VPS and light DevOps work. The Community Edition's unlimited executions pay for the operational overhead quickly at any meaningful volume. For non-technical teams, n8n Cloud at $24/mo Starter is often a better fit than self-hosting.