n8n vs Zapier vs Make: What AI Agents Really Cost (2026)

Since this was written, OpenRouter's model feed lists GPT-5.4 Mini in place of GPT-5 mini; GPT-5.4 Nano in place of GPT-5 nano.

What one AI agent really costs to run on each platform once tasks, credits, executions, tokens and human review are counted

The same AI agent run costs 1 execution on n8n, 7 credits on Make and 26 tasks on Zapier's default settings. That is not a quirk of one workflow. It is the result of three different decisions about what to count, and in 2026 all three vendors rebuilt those decisions around AI. Make moved its whole billing system to credits and now charges its built-in AI by the token. Zapier priced AI steps by model tier, with the most expensive tier as the default, charges two tasks for every MCP tool call, and is folding its separately billed Agents back into tasks. n8n kept charging per workflow run and shipped an AI Agent node, MCP in both directions and human approval for each tool call on every plan, its free self-hosted edition included.

The problem is that none of the price pages tells you what an agent costs. An agent decides at run time how many tools to call and how much text to read, so a price per step, a price per credit and a price per run diverge exactly where agents are most useful. At 10,000 runs a month, our reference support agent costs about $50 on n8n Cloud in platform fees, $72 on Make and $361 on Zapier with your own model key, and the gap grows with every extra tool call. Then come the costs no vendor prints: the model's tokens, the person who approves the agent's work, the person who patches the server, and the cost of moving when a platform shuts down, which two agent-building products are doing over the next five months.

This guide prices all of it from first principles. It builds a cost model of a single agent run, prices that run on every plan of n8n, Zapier and Make using their November 2026 price lists, works through four volumes and three real-world cases, and puts the hidden costs (review time, security patching, failures, lock-in) next to the subscription. It then compares the agent features that decide whether the cheap option is also the safe one, prices six alternatives, and ends with a decision framework you can apply to your own workload.

Contents

1. Why All Three Rebuilt Their AI Billing in Fifteen Months
2. The First-Principles Model: What an Agent Run Actually Costs
3. n8n: One Execution per Run, Tokens on Your Own Bill
4. Zapier: Tasks, Model Tiers and the Two-Task MCP Call
5. Make: Credits, and Why AI Steps Cost More Than One
6. One Agent, Three Bills: Worked Cost Scenarios
7. The Costs That Are Not on the Pricing Page
8. Agent Capabilities Compared: Models, MCP, Approvals and Spend Caps
9. The Alternatives Worth Pricing Before You Commit
10. Outlook: Where Automation Billing Goes Next
11. The Decision Framework: Which Platform for Which Agent

Scorecard: Automation Platforms for AI Agents (November 2026)

Before the detail, here is how nine options compare for running AI agents, scored on the same five criteria with the same evidence standard. Cost uses the reference agent from section 2 (four tool calls per run) at 2,000 runs a month with your own model key wherever that is possible. Pipedream Workflows and OpenAI's Agent Builder are left out because both are shutting down; section 9 covers them.

How to read the criteria. Cost at agent scale (35%) carries the most weight because it is the question this guide answers: the platform's own fee for the reference agent, excluding tokens where you bring your own key. Agent depth and control (25%) covers model choice, MCP and, above all, approval for individual tools, because a cheap agent that cannot be supervised is not cheap. Integrations (15%) and ease for non-developers (15%) capture build and maintenance cost, and data control and exit (10%) captures self-hosting, data residency and what leaving would cost. Change the weights to match your team: a non-technical team that puts ease at 35% will see Zapier and Make rise, and that is a legitimate result.

1. Why All Three Rebuilt Their AI Billing in Fifteen Months

Between August 2025 and November 2026, each of the three platforms changed how it charges for AI, and none of them changed it the same way. Make replaced operations with credits and started pricing its own AI by the token. Zapier put MCP tool calls on the task meter, sold agents by the "activity", then priced AI steps by model tier and began folding agents back into tasks. n8n kept its per-execution price but added the agent features (MCP in both directions, per-tool human approval, first-class agents) that made that price matter. Around them, two well-known agent builders announced their own shutdowns.

The cause is the same in every case. A classic automation's cost to the vendor is roughly fixed per step: a database write, an API call, a log line. An AI step's cost is variable, set by how many tokens a model reads and writes and how many tools it decides to call. Once agents moved from demos into production in 2025, vendors that bundled inference into a flat unit were exposed to costs they could not predict, and vendors that did not bundle it had to decide whether to start. The timeline below shows how quickly that played out.

The market news of the last fortnight sharpens the stakes. Inference keeps getting cheaper at the low end: OpenAI's GPT-6.1 Terra, released October 26, costs $1 per million input tokens, half of its Sol tier. Accountability is shifting onto whoever deploys an agent: Wells Fargo's updated online agreement, effective November 24, says requests made through a third-party AI agent "may be treated as if they came directly from you" - Wells Fargo. And enterprise software is buying the orchestration layer, from SAP's stake in n8n to Workday's purchase of Pipedream.

For a buyer, that combination has a clear implication. Model costs per token are falling, so the platform's meter is becoming a larger share of what an agent costs; responsibility for an agent's actions is becoming yours, so approval and audit features are becoming requirements; and the platforms themselves are consolidating, so exit cost belongs in the comparison. The next section builds the cost model that ties those three together, and then each platform is priced against it.

2. The First-Principles Model: What an Agent Run Actually Costs

Before comparing price pages, it helps to ask what an automation platform actually sells, because the billing unit is just a proxy for that. Strip away the marketing and every one of these products sells three things. The first is connector maintenance: someone has to keep thousands of third-party APIs working as they change, which is expensive, unglamorous work. The second is a runtime: scheduling, retries, queues, credential storage, logs and the guarantee that a run you started at 3 a.m. finishes. The third is an interface that lets people who do not write code assemble the first two. The price is how the vendor recovers the cost of those three things, and the unit (task, credit, execution) is the measurement they chose to recover it by.

For a decade that choice did not matter much, because classic automation is deterministic. A Zap that copies a form submission into a CRM and sends a Slack message runs the same three steps every time, so a per-step price and a per-run price differ only by a constant. AI agents break that symmetry. An agent decides at run time how many tools to call, how many times to retry, and how much text to read. The number of steps is no longer a property of the workflow you designed; it is a property of each input the agent meets. That single fact explains why all three vendors rebuilt their AI billing in the last fifteen months, and why the "cheapest" platform depends on the shape of your agent rather than the sticker price.

2.1 The four layers of the bill

Every agent run on any of these platforms is paid for in four layers, and the price page shows only the first. The platform fee is what Zapier, Make or n8n charge, in their own unit. Inference is what the model provider charges for tokens, either directly to you (bring your own key) or bundled into the platform's unit. Tool costs are the per-call fees of the services the agent uses, such as a search API, a scraping service or a paid data source. And human cost is the build time, the maintenance time and the review time spent approving or correcting what the agent did.

The useful insight is that these layers scale differently with autonomy. A platform that charges per step grows linearly with the number of tool calls. Inference grows faster than linearly, because each step of an agent loop resends the whole conversation so far. Anthropic's tool-use documentation spells out the mechanics: every request is priced on the total input tokens "including in the tools parameter", plus the tool_use and tool_result blocks, plus a tool-use system prompt of 286 tokens on its current models - Anthropic. Its own worked example shows the second request carrying the first request's messages back to the model. A platform that charges per run stays flat no matter how many steps happen inside it.

• Platform fee scales with your vendor's unit: per step, per credit or per run.
• Inference scales with tokens, and tokens compound with every loop.
• Tool costs scale with calls to paid services, such as web search.
• Human cost scales with failure rate and with how many actions need approval.

Put those together and the question "which platform is cheapest" splits in two. For a short, predictable agent (one model call, one or two tools), the platform fee dominates and per-step pricing is fine. For a long, autonomous agent (eight or twelve tool calls with large tool results), inference dominates, and a per-step platform fee stacks a second growing cost on top of the first. The rest of this guide prices both shapes, because a team that runs one usually ends up running the other.

2.2 The reference agent run

To compare like with like, this guide prices one concrete agent on every platform. It is a support-triage agent: a new email arrives, the agent reads it, then calls four tools (look up the customer in the CRM, search the help docs, check the order status, write a draft reply in the helpdesk), and finally a normal workflow step posts a summary to Slack. That is a modest, realistic production agent, the kind small teams actually run, and it is deliberately not a showcase of autonomy.

The token budget is equally modest. Assume 3,000 tokens of system prompt and tool definitions, 1,500 tokens of email and customer context, about 800 tokens returned by each tool, about 200 output tokens per tool request, and a 300-token final answer. Five model calls happen (four that request a tool, one that writes the answer), and because each call carries everything before it, the run consumes about 32,500 input tokens and 1,100 output tokens. The diagram shows why the input side balloons.

Those 32,500 input tokens are the reason model choice matters more than platform choice for anything but the smallest volumes. Priced at today's list rates, the same run costs very different amounts depending on the model behind it. Claude Haiku 5.5 charges $0.10 per million input tokens, and GPT-6 Luna charges $0.10. GPT-6.1 Terra, released October 26 at half the price of OpenAI's Sol tier - BleepingComputer, charges $1. Claude Sonnet 5.5 charges $2, and Claude Opus 5.5 charges $4.

The spread is 40x between the cheapest and the most expensive model for the identical run. The arithmetic for Sonnet 5.5 is 32,500 x $2 per million ($0.065) plus 1,100 x $10 per million ($0.011), so $0.076 a run, or $76 per thousand. Prompt caching narrows it sharply: Anthropic bills cache hits on Sonnet 5.5 and Opus 5.5 at 0.05x the input price, so if the repeated prefix of calls two to five (24,000 of the 32,500 input tokens) is read from cache, the Sonnet run drops to about $0.03, before the small premium Anthropic charges to write the cache. We cover the mechanics in our guide to prompt caching for agents, and the broader menu of levers in our LLM cost-cutting guide.

2.3 Why the unit is a bet on variance

Here is the structural point the rest of the comparison hangs on. A billing unit decides who absorbs the variance of an agent's behaviour. When a vendor charges per step, every extra tool call the agent decides to make lands on your invoice twice: once as tokens, once as a platform unit. When a vendor charges per run, the platform side is fixed and only the token side varies. When a vendor bundles tokens into its own unit (Zapier's Premium tier, Make's built-in AI provider), it absorbs token variance itself and prices in a margin for doing so, usually by charging more per step.

None of these is wrong. They are different bets, and each suits a different agent. If your agent is short and predictable, a per-step price with tokens bundled is simple, and simplicity has real value for a non-technical team. If your agent is long, variable or high-volume, a per-run platform price with tokens on your own bill gives you the most control, because the expensive layer (inference) is then in your hands: you choose the model, cache the prompt and route easy calls to a cheap model, as our guide to model routing shows. That is the lens for reading each platform's pricing in the next three sections.

3. n8n: One Execution per Run, Tokens on Your Own Bill

n8n is the platform whose billing unit was designed, almost by accident, for agents. Its pricing page puts the rule in one sentence: "An execution is a single run of your entire workflow. It doesn't matter how many steps are in the workflow or how much data it processes. It's still a single execution" - n8n. Every plan includes unlimited users, workflows and integrations, and the only thing you buy is a monthly number of executions. When an agent inside a workflow decides to call four tools or fourteen, the platform side of the bill does not move.

The trade-off is that n8n assumes more of its user. It is a technical canvas built around JSON, expressions and nodes, its AI features are thin wrappers over LangChain, and its cheapest serious option (self-hosting) makes you responsible for servers and patches. For a team with one person comfortable at that level, n8n is usually the cheapest platform in this comparison by a wide margin. For a team without that person, the savings are smaller than they look, for reasons section 7 sets out.

3.1 The plans, and what counts as an execution

n8n Cloud has three self-serve tiers. Starter costs €20 a month billed annually (€24 month to month) for 2,500 executions with 5 concurrent runs; Pro costs €50 a month annually for 10,000 executions with 20 concurrent runs, with a larger Pro tier for 50,000 executions; Enterprise is quoted - n8n. The USD prices use the same numbers. Hosted data lives in Frankfurt, which matters for European buyers. Starter and Pro are capped rather than metered: in n8n's words they "are capped and have no overages", so a busy month queues or stops rather than producing a surprise invoice.

Source: n8n pricing page, last updated October 28, 2026. Business is not yet offered on n8n Cloud.

The definition of an execution has three details that matter for agents. Only production executions count, meaning runs started by triggers, schedules or polling; manual test runs do not, and when a workflow calls a sub-workflow "only the parent (top-level) execution counts" - n8n Docs. For n8n's newer standalone agents, "one turn with an agent counts as one execution", where a turn is one message in and one response out - n8n Docs. And the run-time limit (five minutes on Starter) excludes time spent waiting, "for example in a Wait node or for a human approval", so an approval that takes a reviewer an hour does not kill the run.

For early-stage companies there is a cheaper route into the self-hosted Business tier. The Startup plan gives 50% off Business, so €333 a month billed annually for 40,000 executions, to companies with fewer than 20 employees and less than €5 million in total funding - n8n. n8n also removed its old limit on active workflows in August 2025, so every plan now has "unlimited workflows, steps, and users" - n8n Blog.

3.2 Who pays for the tokens

By default, you do, directly. The AI Agent node calls whatever model you connect with your own API key, and the model provider bills you at its list rate, which is why the inference chart in section 2 applies to n8n almost unchanged. That is the source of n8n's control advantage: you choose the model per node, you can turn on the provider's prompt caching, and you can route easy steps to a cheap model and hard ones to a frontier model inside the same workflow.

Since version 2.36.0 (August 2026), n8n Cloud Starter and Pro also offer optional gateway credits, a prepaid balance that lets you call models without your own keys, billed "per request at the rates listed on the service pricing page" - n8n Docs. They are not available on Enterprise Cloud or self-hosted n8n. Separately, each Cloud plan includes Assistant credits for n8n's AI helper that builds workflows from a description: 1,600 a month on Starter and 4,000 or 9,600 on Pro. Those only pay for building; n8n's documentation is explicit that "running a workflow doesn't use Assistant credits, even if n8n Assistant built it" - n8n Docs.

The consequence for budgeting is a clean split. Your n8n bill is a flat fee set by run volume; your model provider's bill is set by tokens; and the two never multiply each other. That separation is what makes n8n's numbers in section 6 look almost implausibly low next to Zapier's, and it is also why an n8n user who picks an expensive model with no caching can still spend far more than a Zapier user on the Premium tier. The platform is cheap; the inference is entirely your responsibility.

3.3 The AI stack: Agent node, MCP and approval for each tool

n8n's agent tooling has matured quickly. The AI Agent node is built on LangChain's JavaScript framework - n8n Docs, takes a chat model, an optional memory (simple in-session memory, or Postgres, Redis, MongoDB and others) and at least one tool, and loops until it has an answer. n8n shipped the MCP Server Trigger and MCP Client Tool in version 1.88.0 in April 2025, so a workflow can both expose itself as a tool to outside agents and call outside MCP servers, over SSE or streamable HTTP but not stdio - n8n Docs. The canvas below is n8n's own illustration of the basic agent shape.

The three sub-nodes in that picture are the whole cost model in miniature. The Chat Model is where your token bill comes from, the Memory decides how much past conversation is resent on every call (and therefore how fast input tokens grow), and each Tool is a call that costs nothing on n8n's meter. The feature that made n8n credible for supervised agents arrived in version 2.6.0 on January 26, 2026: "You can now require explicit human approval before an AI Agent executes specific tools", including MCP Client tools and sub-workflows exposed as tools - n8n Docs. n8n's pricing table lists that approval, the AI Agent node and both MCP nodes as included on every plan, the free Community edition among them.

In August 2026 n8n went a step further and made agents first-class objects that live beside workflows rather than inside them, still in preview at the time of writing. In the official explainer below, n8n's Liam McGarrigle sets out when to reach for one of the new agents and when a workflow with an AI step is the better tool, which is precisely the cost question: an agent turn is one execution, a workflow run is one execution, and the choice between them is about control and reliability, not price.

The rule of thumb in the video is a good one to adopt. Use an agent for tasks "without fixed, predefined steps", where the model genuinely has to decide what to do; use a workflow where the steps are known, and put the AI only where judgement is needed. The cheaper design is almost always the second, not because n8n charges more for agents (it does not) but because a fixed workflow spends fewer tokens and fails in more predictable ways.

3.4 Self-hosting: a free licence and a real server

The Community edition costs nothing and runs anywhere Docker runs. It lacks the collaboration and governance layer (projects, workflow and credential sharing, SSO, Git-based version control, external secrets, log streaming and multi-main high availability), although queue mode for scaling across workers is included - n8n Docs. For a small team, none of the missing features blocks an AI agent. For a company with several builders and an auditor, several of them do, which is what the self-hosted Business plan sells.

The licence is the other thing to read before you rely on the free edition. n8n's Sustainable Use License allows use "only for your own internal business purposes or for non-commercial or personal use" - GitHub, and n8n says plainly "we do not call ourselves open source". Its licence FAQ rules out hosting n8n as a service for clients to build their own workflows, including through an AI agent acting for the user - n8n Docs. Running your own automations, building workflows for clients, and using n8n behind the scenes of your own product are all allowed.

The hardware itself is modest. n8n's documentation says a single small Hetzner CPX11 instance "is enough" for most usage - n8n Docs, and its benchmark reports up to 220 workflow executions per second on a single instance - n8n Docs. At the other end, its official AWS Terraform reference for a small high-availability deployment estimates about $440 a month on demand - n8n Docs. The honest summary is that the server is cheap and the operator is not: the real self-hosting cost is the person who patches it, which section 7 prices.

3.5 The company behind the canvas

n8n's momentum is real and well documented. It raised a $180 million Series C at a $2.5 billion valuation in October 2025, reporting "6x user growth, 10x revenue growth" - n8n Blog. In May 2026 SAP bought into the company through a secondary share sale at $5.2 billion, with a plan to embed n8n in SAP's Joule Studio, and the announcement put n8n at more than 1,400 enterprise customers and 1.7 million monthly active builders - Highland Europe. Its GitHub repository had 209,500 stars on November 4, 2026 - GitHub, and n8n's integrations directory lists more than 2,000 integrations - n8n.

For a buyer, those numbers answer the question that usually sinks open-source-adjacent tools: will this still be maintained in three years? With a strategic investor and a large paying base, the answer is very likely yes, and the self-hostable code is a second insurance policy on top. We covered n8n's fundamentals in more depth in our practical guide to n8n; the rest of this comparison is about what its per-run meter means once you put real agents on it.

4. Zapier: Tasks, Model Tiers and the Two-Task MCP Call

Zapier is the oldest and largest of the three, and its billing logic is the easiest to state: you pay for successful actions. A trigger is free, a filter is free, a formatter step is free, and every step that does something in another app is one task. Zapier says it moves data across 9,000+ apps - Zapier, and its homepage claims more than 3 million businesses as users - Zapier. That breadth is the product. When a reader asks "does Zapier connect to my obscure invoicing tool", the answer is usually yes, which is not true of Make and is even less true of n8n.

The trouble for agent builders is that Zapier's unit counts actions, and an agent is a machine for producing actions you did not plan in advance. In 2026 Zapier responded with three separate meters layered on top of the task: model tiers for AI by Zapier steps, activities for Zapier Agents, and a fixed two tasks per MCP tool call. Each one is defensible on its own. Together they mean the same agent can cost five times more or less depending on which Zapier surface you build it in, and the default settings point at the expensive end.

4.1 The plans and the task ladder

Zapier sells four plans: Free, Professional, Team and Enterprise. Free gives you 100 tasks a month and two-step Zaps only, which means no real agent work. Professional starts at $19.99 a month billed annually for 750 tasks ($29.99 billed monthly), and Team starts at $69 a month annually for 2,000 tasks - Zapier. The plan level and the task tier are chosen separately: you pick a plan, then a monthly task allowance, and the price climbs with the allowance while the cost per task falls. The number of Zaps is unlimited on every plan, so tasks are the only real limit.

What counts as a task is precise and worth memorising, because it decides your bill. Zapier's help center defines a task as "any successful action that runs in Zapier", states that "Zap triggers never use tasks", and adds that testing a step in the editor is free even though the action really runs - Zapier Help. Zapier's built-in tools (Formatter, Paths, Filter, Delay, Looping, Storage, Tables, Forms) do not consume tasks either. One shared pool now covers Zaps, AI by Zapier, Code by Zapier and MCP, so every surface drains the same allowance.

Prices per month in USD, from Zapier's pricing page as of November 4, 2026. Tiers between those shown exist.

Two details on the overage policy change the real cost more than the headline. Pay-per-task overage is billed at 1.25x the base rate on annual plans but 2.5x on monthly plans, and even with overage switched on, usage stops at 3x your subscription until the next cycle - Zapier. The help center adds that pay-per-task is on by default for accounts created after January 2024 - Zapier Help. For an agent whose volume you cannot predict, this matters: a busy month on a monthly plan costs two and a half times the advertised rate for every task above the line, which is the opposite of the volume discount the ladder implies.

4.2 AI by Zapier: the model tier multiplies everything

The change that matters most for AI agents arrived on June 15, 2026, when Zapier moved AI by Zapier steps to model-tier pricing. The formula, from Zapier's own help article, is: tasks per run = (1 x model rate) + (number of tool calls x model rate) - Zapier Help. The model rate is 1x for Standard models, 3x for Advanced and 5x for Premium, and bringing your own API key costs 1x but works only with Advanced or Premium models. A Premium step that calls two tools therefore costs 15 tasks, not 3.

Three facts on that page deserve attention from anyone budgeting. First, a new AI by Zapier step on a paid plan defaults to Premium, the most expensive tier. Second, the Auto setting picks Claude 4.5 Haiku on Advanced and Claude 4.6 Sonnet on Premium, which are not the newest models their maker sells. Third, a single step stops at 75 tasks by default and asks for approval before continuing, a ceiling you can raise to 500. That last one is a genuine safety feature: it is a per-run spending cap, which is exactly what an agent needs.

Here is what the formula does to a realistic agent. Take the reference run this guide uses throughout: an agent that reads a support email, calls four tools (look up the customer, search the help docs, check the order, draft the reply), then posts a summary to Slack. On the Premium default that is (1 + 4) x 5 + 1 = 26 tasks per run. With your own key it is (1 + 4) x 1 + 1 = 6 tasks per run, plus whatever your model provider charges you for the tokens. The screenshot below is the selector where that choice is made.

The dropdown makes the trade explicit, and it is a fair trade in one sense: Premium includes the tokens. If you would otherwise pay a frontier model's list price, a 5x task multiplier can be the cheaper way to buy inference at high volume, a point the worked scenarios in section 6 return to. What you give up is choice. You get the model Zapier picked for the tier, on Zapier's schedule, and you cannot apply your own prompt caching or route easy calls to a cheaper model.

4.3 Zapier Agents and the activity meter

Zapier Agents, the standalone autonomous assistants, run on a different unit entirely: activities. An activity is "any billable action that your agent takes", which includes using a trigger, answering from a knowledge source, taking an action, browsing or scraping a page, or running a web search - Zapier Help. Activities do not touch your task pool. The Free plan includes 400 activities a month, the paid Agents Pro add-on includes 1,500, and a single run is capped at 10 activities on Free and 40 on paid plans. When the allowance runs out, agents stop running, though you can still chat with them.

Agents Pro costs $400 a year ($33.33 a month), or $50 billed monthly, on top of any Zapier plan - Zapier. At 1,500 activities that is roughly 300 runs of our five-activity reference agent, which is a pilot budget, not a production one. The bigger news is that Zapier is retiring the separate meter. A help article published in July 2026 says Zapier "is migrating Agents (agents.zapier.com) to AI by Zapier" and that "you no longer manage a separate Agents subscription or add-on fees", while noting that no shutdown date has been set - Zapier Help.

So the brief version of 2026's story is that Zapier briefly sold agents by the activity and is now folding them back into tasks, priced by model tier. That is a sign the company found the separate unit confusing for buyers, and it means any long-term cost plan for Zapier agents should be built on the AI by Zapier formula above rather than on the activity allowance.

4.4 Zapier MCP: two tasks per tool call

Zapier MCP turns Zapier's catalogue into tools that an outside agent can call: Claude, ChatGPT, Cursor or your own code connect to a Zapier-hosted MCP server and use Zapier's integrations as if they were native tools. The billing is a flat rule: "Each successful tool call through your MCP server consumes two tasks", failed calls are free, and listing tools, authenticating and viewing history do not count - Zapier Docs. Zapier's own example: "Add 5 rows to a spreadsheet" is five tool calls and ten tasks.

MCP moved onto the task meter on September 18, 2025, when Zapier's blog announced that MCP was now included in the Free, Pro and Team plans and its comparison table read "1 call = 2 tasks" - Zapier. Two adjustments arrive on December 1, 2026: tool calls to free by-Zapier apps such as Tables and Forms stop counting, and AI by Zapier steps called through MCP are charged at their own model-tier price. The same date also starts billing for Next Gen Zaps and the Zapier SDK, with each successful Next Gen run using at least one task - Zapier Help.

Zapier's own workshop shows the setup end to end: installing Zapier MCP inside Claude, connecting two apps and running a real action from one prompt. It is worth watching before you price MCP, because the demo makes it obvious how quickly one conversational request turns into several tool calls.

The economics of MCP are clearest in our reference run. If the agent lives in Claude or your own code and uses Zapier only for its four tool calls, the run costs 8 tasks on Zapier plus your own token bill. That is more than the 6 tasks of an AI by Zapier step on your own key, because MCP charges two tasks per call where the AI step charges one at the 1x rate. MCP's advantage is not price but placement: the reasoning happens in the agent you already use, and Zapier becomes the connector layer. The usage sidebar below is where those MCP tasks show up, separately from your Zap usage.

Watch that split closely. A Zapier community thread documented a user on a 100-task plan who was told they had exceeded their limit after seeing only 25 tasks in the usage graph, because the graph showed Zap usage while the MCP calls were counted elsewhere - Zapier Community. The lesson generalises: when an agent consumes a meter, check where the meter is displayed before you trust a dashboard.

4.5 Where Zapier earns its premium

Zapier is the most expensive way to run high-volume agents, and it is still the right choice for a specific buyer. If your team is non-technical, your apps are numerous and unusual, and your volume is hundreds rather than tens of thousands of runs a month, Zapier's catalogue and polish save more in setup time than they cost in tasks. Its Human in the Loop app pauses a Zap until a person approves or adds data, and only the successful step counts as a task - Zapier Help. AI by Zapier can also require approval before it uses any individual tool, which is the per-tool gate a careful agent deployment needs.

The building experience is cheap to explore as well. Zapier Copilot, the assistant that drafts Zaps from a description, does not use tasks while you build and test; only live runs count - Zapier Help. For enterprises, Zapier can route AI processing through the customer's own AWS Bedrock account, which addresses the data-control objection that usually pushes regulated buyers toward self-hosting. The cost of all this convenience is the multiplier: once runs reach the thousands, the per-step meter is the dominant number on the invoice, and the next two sections show how differently Make and n8n treat that same run.

5. Make: Credits, and Why AI Steps Cost More Than One

Make (formerly Integromat, part of process-mining company Celonis since 2020 - TechCrunch) sits between Zapier and n8n on almost every axis. Its visual canvas is friendlier than n8n's and more powerful than Zapier's, its catalogue lists 3,500+ apps - Make, and its price per unit is the lowest of the three for small and mid-sized volumes. It meters each module that runs, like Zapier, but its unit is much cheaper, and since 2025 it charges for its own AI by the token rather than by the step.

That last change is the one buyers noticed. On August 27, 2025, Make replaced operations with credits as its billing unit, converting balances "at a 1:1 ratio" - Make Help. Make's position is that nothing changed for normal automations: "1 operation still uses 1 credit", while usage "may increase if you're using Make's AI Provider, as we align billing with actual AI usage costs" - Make Help. Both halves of that sentence are true, and the second half is where agents live.

5.1 What a normal scenario costs in credits

Make's counting rules are more granular than Zapier's, and each rule has a cost consequence. A trigger module uses 1 credit per run "regardless of whether they return data. Even when a trigger finds no new data, it still uses 1 credit for checking" - Make Help. A search module uses 1 credit per run however many rows it returns, an action module uses 1 credit per bundle it processes, and an iterator costs a credit to split an array and then a credit per item in every module after it. Routers, filters and error handlers are free.

The trigger rule is the one that catches people. A scenario scheduled to check for new data every minute runs about 43,200 times in a 30-day month, and each check costs a credit even when nothing has arrived, which is more than four times the 10,000 credits of an entry plan spent on an idle scenario. Zapier does not charge for polling at all. The fix on Make is design, not plan: use instant (webhook) triggers where the app supports them, or schedule checks at the interval the business actually needs. The iterator rule explains the angriest reactions to the credits launch, such as the user who pointed out that "an automation that modifies 10 lines of a spreadsheet counts as 10 credits" and called the change "a brutal price increase" - Make Community.

5.2 The plans, and a price that depends on your country

Make sells Free, Core, Pro, Teams and Enterprise. Free includes 1,000 credits a month, two active scenarios and a 15-minute minimum interval. The pricing page itself shows Core at $9 a month billed annually for 10,000 credits, Pro at $16 and Teams at $29 - Make. But the prices the page actually displays are loaded per country: querying the price feed behind the page on November 4, 2026 returned $12 a month (annual) or $16 (monthly) for Core at 10,000 credits for US visitors, $11 annual for Germany, the UK and France, and the $9 figure for many other countries. Every Make price in this guide is the US figure, so readers elsewhere may pay less.

US prices per month, from the price feed behind make.com/pricing on November 4, 2026. Core tops out at 300,000 credits a month; Pro scales to 8 million.

Three policies change the effective price. Extra credits, bought by hand or automatically, carry "a 25% additional cost compared to the credits in your plan" - Make Help, and automatic top-ups are capped at your plan's own credit amount per cycle. Unused credits do not roll over: "Credits expire at the end of the term". And on an annual Core plan you still receive a monthly allocation; only annual Pro and Teams plans get the year's credits as one pool - Make Help, which is worth real money if your agent volume is seasonal. Since November 2025 every paid plan can connect its own AI provider key, a feature that used to require Pro - Make Help.

5.3 Make's AI Provider: credits by the token

When you use Make's AI Provider instead of your own key, credits "vary based on the input and output tokens consumed and the model selected" - Make Help. Make publishes the exchange rate for each model as tokens per credit. Its own tiers run on OpenAI's small models: Small and Medium use GPT-5 nano at 18,080 input and 2,260 output tokens per credit, and Large uses GPT-5 mini at 3,616 and 452. Premium models cost far more credits: Make lists Claude Sonnet 5 at 452 input and 90 output tokens per credit, GPT-5.6 Terra at 452 and 75, and Claude Opus 5 at 180 and 36.

The exchange rate lets you check whether Make marks inference up, and the answer is "roughly not". At the entry Core price of $12 for 10,000 credits, a credit costs $0.0012, so Sonnet 5 input through Make works out at about $2.65 per million tokens against Anthropic's list price of $2 for Sonnet 5 - Anthropic. On the 300,000-credit Core tier the same tokens cost about $1.75 per million, below list. Make also cut AI credit usage on August 25, 2026 by pricing input and output tokens separately, and its example step fell from about 0.6 to 0.16 credits, "a 73% reduction" - Make Help.

Fixed-price AI features are the exception to keep an eye on. The AI Content Extractor charges 10 credits per document page or invoice and 20 credits per minute of transcribed audio, and the Make Code app consumes "two credits for every one second of code execution time" - Make Help. These are the modules where a credit is no longer a cheap unit, and they are where the "AI features cost more than one credit" headline comes from. A document-heavy agent that extracts 50-page contracts spends 500 credits per document before it reasons about anything.

5.4 Make AI Agents: one credit per run, plus every tool

Make's rebuilt agent product, Make AI Agent (New), was released on February 2, 2026 and is still labelled an open beta, with Make warning that "product functionality and pricing may change" - Make Help. Its billing is set out in a table: running an agent costs "1 credit per operation" with your own key, plus "credits based on AI tokens" on Make's AI Provider, and each tool the agent calls adds its own operations - Make Help. The reasoning panel below, from Make's documentation, shows one agent run with two tool calls billed at 25.64 credits for "1 operation".

The gap between "1 operation" and "25.64 credits" is the AI Provider's token charge, and it is the number to watch. For our reference run, the arithmetic is: 1 credit for the trigger, 1 for the agent run, 1 for each of four tool operations and 1 for the Slack post, so 7 credits with your own key. On Make's Large tier the tokens add about 11 credits, for roughly 18 per run. On Claude Sonnet 5 through Make's provider the tokens add about 84 credits, for roughly 91 per run, which is when Make stops being the cheap option.

Make has been adding cost controls to its agents at a fast pace. In July 2026 it added sub-agents and tool output filtering, which trims what a tool returns before it reaches the model, and by October it had added prompt caching so an agent can reuse its instructions and tool definitions, "reducing latency and credit usage" - Make Help. Both attack the input-token growth from section 2 directly. Make's own walkthrough of the caching feature is a useful twelve minutes if you plan to use its AI Provider.

Two gaps remain. The Make MCP server is included on all plans and lets an outside agent run your scenarios as tools, with management tools reserved for paid plans - Make Help; the scenarios it runs bill like any other run. And human approval is not yet a general feature: at its Waves '26 event on October 20, Make listed "Human in the loop (HITL): Pause any scenario for approval and sign-off from Slack, Teams, or the Make mobile app" as coming soon - Make. Until it ships, approval on Make means building your own pause-and-resume pattern.

5.5 Maia, Grid and where Make wins

Make's AI assistant, Maia, builds scenarios from a conversation, and unlike Zapier's Copilot it is not free to talk to: every message costs credits, Make's own example reply used 140.04 credits, and paid plans get 70 free messages within 30 days - Make Help. That is cheap compared with an hour of a consultant's time and expensive compared with a free assistant, so it is worth knowing before you spend an afternoon iterating with it. Make Grid, its map of every automation in an account, can now also display n8n workflows, Claude Managed Agents and Relevance AI assets - Make Help, a sign that Make expects to be one of several platforms in a company rather than the only one.

Make wins clearly in one zone: visual builders with low to moderate volume who bring their own model key. At 2,000 reference runs a month it costs about the same as n8n Cloud, needs no server, and offers a gentler interface than n8n with a much cheaper unit than Zapier. It loses ground in two situations: agents that call many tools (each one is a credit, where n8n charges nothing) and agents that run on Make's AI Provider with a premium model (where the token credits dominate). The next section puts all three side by side so those zones are visible.

6. One Agent, Three Bills: Worked Cost Scenarios

The previous three sections described each meter on its own terms. This one runs the same agent through all of them, at four volumes, so the differences stop being abstract. The method is deliberately simple and reproducible: the reference run from section 2 (a trigger, an agent that makes four tool calls, and a Slack post), US list prices, annual billing, and for each volume the cheapest plan that covers it, including each vendor's own overage or extra-credit rules. Token costs are shown separately, because on every bring-your-own-key option they go to your model provider, not to the platform.

Per run, the reference agent consumes 6 tasks on Zapier with your own model key (one AI by Zapier step at 1x plus four tool calls, plus the Slack step), 8 tasks if an outside agent calls Zapier through MCP, 7 credits on Make with your own key, and 1 execution on n8n. Those four numbers, multiplied through each vendor's price ladder, produce everything below. If your agent makes more or fewer tool calls, the last subsection shows how each meter responds.

6.1 The platform fee at four volumes

The first comparison isolates the platform fee, with tokens excluded on every row because all of these options send the model bill to your provider. It answers the narrow question "what does the automation platform itself charge to run this agent?", which is the number most comparison articles are actually arguing about.

US list prices, annual billing, cheapest plan per volume including overage or extra credits. n8n prices are the euro figures, which n8n also charges in dollars. Tokens excluded.

Three patterns stand out. At low volume, Make is the cheapest platform, because its 10,000-credit entry tier costs less than n8n's Starter. Around 2,000 runs the two are level. From there n8n pulls away, because its unit does not grow with the four tool calls inside each run, and at 50,000 runs it costs less than half of Make and roughly a ninth of Zapier. Zapier is the most expensive at every volume, and the gap widens with scale, which is the per-step meter doing exactly what section 2 predicted.

6.2 The all-in bill, with tokens included

The platform fee is only half the bill, and the bundled options (Zapier's Premium default, Make's AI Provider) include tokens that the other rows exclude. To compare honestly, this table adds inference for a Sonnet-class model. The bring-your-own-key rows use Claude Sonnet 5.5 with prompt caching on the repeated prefix, about $0.03 a run; the bundled rows use the model each platform provides at a similar level (Zapier's Premium Auto choice of Claude Sonnet 4.6, and Claude Sonnet 5 on Make's AI Provider).

The bundled options are not the rip-off they first appear. At Anthropic's list price for Sonnet 4.6 ($3 in, $15 out per million tokens), the reference run's tokens alone would cost about $0.11, and Zapier charges 25 AI tasks for them, about $0.14 at the 50,000-task tier, so the Premium multiplier is close to the cost of the model it buys. The expense comes from what bundling prevents: you cannot cache the prompt, you cannot choose a newer or cheaper model, and the platform's per-step fee is stacked on top. Bring your own key, enable caching, and the same agent costs roughly a third as much on n8n or Make as on Zapier's default.

6.3 What happens as the agent makes more tool calls

The reference agent makes four tool calls, but agents vary. This chart holds volume and model fixed and changes only the number of tool calls per run, pricing each platform's unit at a mid-size plan's rate: $4.89 per 1,000 tasks on Zapier (the 100,000-task tier), $0.90 per 1,000 credits on Make (the 80,000-credit Core tier) and $5.00 per 1,000 executions on n8n (Pro, 10,000 executions). It shows the platform fee only.

The crossover between Make and n8n sits between two and four tool calls: below it, Make's cheap credit wins; above it, n8n's flat execution wins, and the gap grows with every call. Zapier's line is steeper than both at every point. Keep the token side in mind too, because it grows faster than any of these lines: on Sonnet 5.5 without caching, the same run costs about $25 per 1,000 runs at one tool call and about $300 at twelve, since every loop resends a longer conversation. For autonomous agents, the platform fee is the smaller and more predictable number.

6.4 Three realistic cases

A solo consultant runs a lead-qualification agent 300 times a month, with three tool calls per run. Zapier on its own key costs $39 a month (and about $118 on the Premium default), Make costs $12, n8n Cloud costs $20, and inference on Claude Haiku 5.5 costs under a dollar. At this scale the platform fee barely matters, and the right choice is whichever tool the consultant can build and fix fastest. If that is Zapier, the extra $27 a month is cheaper than an afternoon of learning n8n.

A 20-person software company runs a support-triage agent 2,000 to 10,000 times a month, the reference case. Here the choice starts to matter: between $80 and $380 a month all-in on n8n or Make with caching, against $300 to $1,060 on Zapier's default settings. The decision hinges less on price than on who maintains it. If an engineer owns automations, n8n Cloud is the efficient answer; if an operations manager does, Make offers most of the savings with a gentler canvas.

A data team runs a lead-enrichment agent 50,000 times a month with eight tool calls each. Platform fees are $120 on n8n Pro, about $484 on Make and about $1,499 on Zapier with its own key, and inference on Haiku 5.5 adds about $430 on every platform. At this volume self-hosted n8n also becomes worth pricing seriously, since the licence costs nothing and a small server handles the load, as long as someone owns patching. These three cases sketch the decision rules that section 11 formalises.

7. The Costs That Are Not on the Pricing Page

Every number so far has been a subscription or a token price, because those are the numbers vendors publish. They are also the smaller part of the bill for most teams. The real cost of an AI agent in production includes the people who build, maintain and supervise it, the security work of keeping the runtime patched, the money lost when an agent does the wrong thing, and the cost of moving when a platform changes under you. None of these appear on a pricing page, and all of them differ sharply between Zapier, Make and n8n.

This section prices them as concretely as the evidence allows. Some figures are assumptions you should replace with your own (your hourly cost, your approval rate), and they are labelled as such. The point is not the exact total but the ranking: in most real deployments, at least one of these hidden lines is larger than the platform fee, and choosing a platform on the fee alone optimises the wrong number.

7.1 Human review: the line that usually wins

The most underestimated cost of an AI agent is the person approving it. Suppose every one of our 2,000 monthly reference runs needs a human to approve the drafted reply, and each approval takes 30 seconds. That is 1,000 minutes, or about 16.7 hours a month. At an assumed loaded cost of $40 an hour (use your own figure), review costs about $667 a month, which is more than the Zapier Premium bill for the same runs and more than twenty times the n8n bill. Human-in-the-loop is not free just because the approval node is.

The lever is design, not platform. If you gate only the risky tool (sending the reply) and let the agent run its read-only lookups freely, and if after a month of evidence you auto-approve the 80% of cases the agent gets right, review falls to about $133 a month. This is why per-tool approval matters so much more than a generic "pause the workflow" step. n8n's AI Agent can require approval for "all tools connected to an AI Agent node, or just selected individual tools", with nine channels for the reviewer including Slack, Teams, Gmail and its own chat - n8n Docs. AI by Zapier can require approval before any individual tool runs. Make's approach is covered in section 5.

• Gate writes, not reads: approvals on lookups waste reviewer time.
• Measure before relaxing: auto-approve only categories with a track record.
• Route to where reviewers are: an approval in Slack gets answered faster than one in a dashboard.

The practical implication is that the cheapest platform for a supervised agent is the one whose approval design lets you put a human exactly where the risk is and nowhere else. Reviewer time also behaves differently from every other cost in this guide: it does not get cheaper with volume, and it does not fall when models get cheaper. Treat reviewer minutes per run as a first-class metric, track it next to cost per run, and you will find that a two-second approval in a chat thread beats a cheaper platform with a clumsy approval page every month.

7.2 Security and patching: what self-hosting really asks of you

Self-hosted n8n is the cheapest option on paper because the software is free, and it is the most expensive one in a specific way: you are the security team. 2026 made this concrete. In January the US National Vulnerability Database published CVE-2026-21858, rated 10.0 Critical, under which n8n versions from 1.65.0 to before 1.121.0 let an unauthenticated remote attacker access files on the server through certain form-based workflows - NVD. In February came CVE-2026-27493, a 9.0-rated expression injection in Form nodes that could escalate to remote code execution when chained with a sandbox escape - NVD.

Two later advisories are specific to AI agents and worth knowing by heart. CVE-2026-56349 describes a Guardrail-node bypass in n8n before 2.10.0 that let end users "craft malicious inputs to circumvent guardrail protections", which is prompt injection by another name - NVD. And CVE-2026-42228, published in May, found that the hosted chat endpoint did not check whether a connection was authorised to interact with an execution, so an attacker who could identify the ID of a workflow in a waiting state could attach to it and receive the prompt meant for the legitimate user - NVD. A workflow waiting for a human is exactly what an approval step creates.

None of this means n8n is unsafe; it means the project is popular, scrutinised and patched quickly, and every fix was available when the advisory was published. The cost question is who applies the fix. On n8n Cloud, Zapier and Make, the vendor patches for you. On self-hosted n8n, someone on your team has to notice the advisory, upgrade and retest, every time. Budget for it explicitly: a few hours a month for a single small instance is a common estimate, and more once you run queue mode across several workers. Our guides to prompt injection defence and agent sandbox controls cover the controls that limit the damage when a patch is late.

7.3 The cost of a wrong action, and the new bank rules

An agent that sends the wrong email costs an apology. An agent that moves money, deletes records or changes a customer's account can cost far more, and in 2026 the institutions on the other side of those actions started saying so in writing. Wells Fargo's updated Online Access Agreement, effective November 24, 2026, says that if you use a third-party AI tool or agent to access your accounts, "requests submitted through the tool may be treated as if they came directly from you", and that the bank reserves "the right to restrict or block access from certain third-party AI tools" - Wells Fargo.

For automation builders the lesson is that liability for an agent's action now sits, explicitly, with whoever deployed it. That changes the expected cost of an unsupervised agent with write access to financial systems from "the platform fee" to "the platform fee plus the probability of a bad action times its cost". The mitigation is structural: keep agents that touch money behind per-tool approval, give them the narrowest credentials the task needs, and log every tool call with its inputs. Our guide to non-human identity for agents covers how to scope those credentials.

7.4 Failures, retries and test runs

Each platform counts failure differently, and the difference compounds for agents, which fail more often than deterministic workflows. Zapier is the most forgiving: a task is only a successful action, failed MCP calls are free, and testing in the editor does not use tasks. Make counts operations as they execute, so a scenario that fails halfway has still consumed the credits of the modules that ran. n8n counts an execution per workflow run, and whether a failed or manual run counts depends on the plan and how it was triggered, which section 3 covers.

The way to turn this into money is to measure your agent's failure and retry rate in a pilot, then multiply. An agent that retries 20% of its runs once costs 20% more on any per-run or per-step meter, and its retries also re-spend tokens. If your error rate is high, Zapier's "only success counts" rule is worth real money; if your error rate is low, the difference is noise. Either way, observability pays for itself here, which is why we compared the main tools in our review of agent observability platforms.

7.5 Lock-in and migration

The last hidden cost is the one you pay when you leave. Workflows built in any of these tools are stored in the vendor's own format, and an agent's prompts, tools, memory and approval logic are woven through them. Moving 40 production automations is weeks of work: rebuild, retest, re-approve, rerun in parallel. Section 9 described two products that are forcing that work on their users right now, with Pipedream Workflows closing in March 2027 and OpenAI's Agent Builder in November 2026.

Lock-in differs by platform. Zapier and Make are cloud-only, so leaving means rebuilding. n8n workflows are JSON files you can export, version and run on your own server indefinitely, which is the strongest exit option of the three, although moving off n8n to something else is still a rebuild. The cheapest insurance is architectural: keep the agent's reasoning layer (prompts, tool definitions, model choice) portable, expose your own systems through standard MCP servers you control, and treat the automation platform as replaceable plumbing. Our guide to building a remote MCP server shows how small that investment can be.

8. Agent Capabilities Compared: Models, MCP, Approvals and Spend Caps

Cost per run only matters if the agent works, and the three platforms differ as much in what they let an agent do safely as in what they charge. For AI agents specifically, four capabilities decide both reliability and cost: which models you can use (and whether you can cache and route them), which direction MCP works in, how precisely a human can approve an action, and what stops a runaway agent from spending your month's budget in an afternoon. Integration count, the traditional headline, matters less than it used to, because MCP lets an agent reach tools that its platform never built a connector for.

The table summarises where each platform stood on November 4, 2026. Every cell comes from the vendor's own documentation cited in sections 3 to 5, and features marked beta or coming soon are labelled as such, because a feature in beta has a way of changing price when it leaves beta.

8.1 Models: choice is a cost lever, not a feature checkbox

The model column matters more than it looks. On n8n, and on Make or Zapier with your own key, you pick the model per step, so you can send classification to a model like Claude Haiku 5.5 at $0.10 per million input tokens and reserve a frontier model for the one step that needs judgement. That routing pattern can cut inference bills by more than half, as our guide to model routing for agents shows with worked numbers. On a bundled tier you get one model per step, chosen by the platform.

Bundled models also lag. Zapier's Premium Auto setting picks Claude Sonnet 4.6 and Make's AI Provider lists Claude Sonnet 5, while Anthropic's current mid-size model is Claude Sonnet 5.5 at $2 per million input tokens, the same input price as Sonnet 5 with cheaper cache hits. A platform's model list is updated on the platform's schedule, and an agent locked to an older model pays more for less capability for as long as the list stays behind. Our benchmark and pricing tracker is a quick way to check what a newer model would save.

8.2 MCP: which way the tools flow

MCP support sounds identical across the three, but the direction differs, and direction determines the cost model. Zapier MCP is a server for outside agents: an agent living in Claude, ChatGPT or Cursor calls Zapier's catalogue, and Zapier charges two tasks per successful call. Make's MCP server lets an outside agent run your scenarios, which bill like any scenario run. n8n works both ways: its MCP Server Trigger exposes a workflow as a tool that outside agents can call, and its MCP Client lets an n8n agent call external MCP servers. The screenshot below, from n8n's release notes, shows a single MCP Server Trigger exposing four tools at once.

The practical upshot is a design choice with a price attached. If your team already lives in Claude or ChatGPT and wants those assistants to act in business systems, the reasoning (and its token bill) sits with your AI subscription, and the automation platform is billed per tool call. On n8n, a tool call that an outside agent makes into one of your workflows is one execution of that workflow, so a whole sub-process (look up, validate, write, notify) costs the same as a single step. For a deeper look at how to expose your own systems this way, see our guide to LLM tool gateways.

8.3 Approvals and spend caps: the safety features with a price

Approval design was priced in section 7.1, and the capability gap is real. n8n has had per-tool approval on every edition since January, Zapier offers per-tool approval inside AI by Zapier plus a paid Human in the Loop app, and Make has announced approvals but not shipped them. For a supervised agent that touches customers or money, that gap alone can decide the platform, whatever the price per credit.

Spend caps are the less glamorous twin, and they work differently on each platform. n8n's capped Starter and Pro plans cannot produce an overage at all, which suits a team that would rather have an agent pause than an invoice surprise. Zapier's 75-task default stop per AI step is a true per-run cap, the best defence against an agent stuck in a loop, but its account-level pay-per-task setting is on by default for newer accounts and bills monthly plans at 2.5x. Make caps automatic top-ups at your plan's own credit amount. Whichever you choose, set the cap deliberately on day one, because an agent's worst month is always the one you did not budget for.

9. The Alternatives Worth Pricing Before You Commit

The three incumbents are not the only places to run an agent, and a cost comparison that stops at them misses the most interesting design choices of 2026. A wave of newer products started from the agent rather than from the workflow, and their billing units reflect it: some charge per run like n8n, some pass model costs through at list price, and some charge per action like Zapier. Looking at them is useful even if you never switch, because each one shows a different answer to the variance question from section 2, and a good answer elsewhere is leverage when you negotiate with your incumbent.

There is also a darker reason to look around. Two well-known agent-building products are closing in the next five months, which is the most concrete evidence available that the platform you choose is itself a cost. The subsections below cover the run-priced open-source option, the agent-first products, the action-priced enterprise options, the shutdowns, and the path for teams that would rather not build workflows at all.

9.1 Activepieces: run-based pricing with an MIT core

Activepieces is the closest structural cousin to n8n, with one important licensing difference. Its core is released under the MIT license, with enterprise and cloud features under a commercial license - Activepieces, so a company can self-host and modify the core without the "internal use" restrictions of n8n's Sustainable Use License. The project had about 25,000 GitHub stars on November 4, 2026 - GitHub, roughly an eighth of n8n's count, which says a lot about the size of the community you would be relying on for templates and answers.

Its pricing reads like a deliberate blend of n8n and Make. "A run is one credit however many steps it has", but AI sits on top: a step on a fast model costs 2 credits, a smart model 10 and a frontier model 20, agentic actions cost 1 credit each, and connecting your own model keys on paid plans "drops every AI step to 1 credit" - Activepieces. The Plus plan is $20 a month for 10,000 credits, Team is $200 for 50,000, and overage on both is $0.007 per credit. On our reading of those rules, the reference agent costs about 15 credits per run on a smart model, or about 6 with your own key, so 2,000 runs a month land near $34 on Plus with your own key plus your token bill.

• Licence: MIT core, commercial enterprise layer, free self-hosting.
• Agents: built-in MCP server; approvals pause a flow via a resume URL.
• Watch for: a pricing page that contradicts itself on whether free credits reset daily or monthly.

Activepieces is a sensible choice for a technical team that wants n8n's per-run economics with a more permissive licence, and it is a credible threat in any negotiation with n8n. What it lacks is the depth of n8n's AI tooling (we found no per-tool approval gate for agent tool calls, only a flow-level approval step) and the size of n8n's ecosystem. For a small team, that ecosystem is worth more than the licence, because an obscure integration that already exists as a community node is a day of work you do not do.

9.2 Gumloop and Lindy: agent-first and priced through credits

Gumloop takes the most transparent position on inference of any product reviewed here. Its Pro plan "starts at $37/month" with 20,000 credits included, a credit is "indexed at $0.005", and AI usage is billed at the provider's list price plus an 8% orchestration fee, rising to 16% if you bring your own model key - Gumloop. In plain terms, Gumloop is a pass-through for tokens with a stated margin, which is the honest version of what every bundled-AI platform does. It has a per-connector approval setting ("Ask for writes/deletes") and hosts MCP servers. The company raised a $50 million Series B led by Benchmark in March 2026 - SiliconANGLE.

Lindy has moved further toward being an assistant than a workflow tool, selling itself as a teammate that lives in Slack. Its Team plan starts at $29.99 per active user a month for 3,000 credits, with top-ups at $10 per 1,000 credits - Lindy, and credits cover model use as well as actions, with typical everyday requests costing 2 to 250 credits. Lindy does not bill overage automatically; it pauses until the next cycle or a top-up, which is a safer default than Zapier's pay-per-task. Its built-in rule that "anything with outside impact waits for approval" is the strictest approval default of any product here.

The cost logic of both is different from the incumbents in a way worth naming. They are priced for agent work measured in outcomes, not for integration work measured in steps. That makes them good value for knowledge-work agents (research, inbox triage, meeting prep) whose steps are mostly model calls, and poor value for high-volume plumbing (syncing 50,000 records a day), where a per-run platform such as n8n is far cheaper. If your agent mostly thinks, look here; if it mostly moves data, stay with the workflow tools.

9.3 Relevance AI and Copilot Studio: priced per action

Relevance AI has split its bill into two meters. An Action is "a single run of a Tool", and it counts even if the tool fails; Vendor Credits pay for the model "with no markup", and bringing your own key is optional on every plan - Relevance AI. Its documentation still lists a Pro plan from $19 a month billed annually with 2,500 actions, although the public pricing page now leads with enterprise sales, and extra actions cost $80 per 1,000. Per-tool approval modes ("Approval Required" or "Auto Run") are built in.

Microsoft Copilot Studio is the one to price if your company already pays for Microsoft 365. Capacity packs cost $200 a month for 25,000 Copilot Credits - Microsoft, and Microsoft's billing table charges 5 credits per agent action and 2 per generative answer, while internal use by people who hold a Microsoft 365 Copilot licence is listed as no charge - Microsoft Learn. At the pay-as-you-go rate of a cent per credit, our reference run's four tool actions and one answer cost about $0.22, which is in Zapier Premium territory. Microsoft told investors on October 28 that customers consuming credits through usage-based billing grew more than 75% in the quarter - Microsoft.

Both products confirm the pattern from section 2: when the unit is the action, the bill scales with how many things the agent does, and the vendor's margin rides on autonomy. That is acceptable when actions are valuable and rare (an agent that updates an ERP record after a long analysis) and painful when they are cheap and frequent. The Microsoft case has one extra wrinkle worth checking with your administrator: "no charge for licensed users" covers internal agents only, and publishing to an external channel needs the standalone licence.

9.4 The shutdown warning: Pipedream Workflows and OpenAI Agent Builder

The most important alternative facts of the year are not prices. Pipedream, the developer-friendly workflow platform that Workday agreed to buy in November 2025 - SiliconANGLE, now says that "Pipedream Workflows and String will shut down on March 31, 2027", that Workflows are in maintenance mode, and that it is no longer accepting signups - Pipedream. What survives is Pipedream Connect, an integration API for apps and AI agents, priced at $99 a month billed annually with 10,000 credits and $0.012 per extra credit, where a credit is 30 seconds of compute - Pipedream.

OpenAI's Agent Builder, the drag-and-drop canvas that launched with AgentKit at DevDay in October 2025 and was widely read as a threat to n8n, is closing too. OpenAI's deprecations page says it notified developers on June 3, 2026 and that "Agent Builder is scheduled to shut down" on November 30, 2026, with migration paths to the Agents SDK or ChatGPT workspace agents - OpenAI. That is less than four weeks from the date of this guide.

Neither shutdown is a scandal; products change. But they put a number on a cost that comparison articles ignore: migration. Every workflow built on a closing platform has to be rebuilt, retested and re-approved, and that labour usually exceeds years of subscription fees. It is the strongest argument for n8n's self-hosted option (the code keeps running even if the company changes course) and for keeping agent logic in portable form, such as prompts and tool definitions you own, behind standard interfaces like MCP. Our explainer on MCP and A2A covers why those standards matter for exactly this reason.

9.5 When you do not want to build workflows at all

Everything above assumes you want to assemble automations yourself, step by step, and own their upkeep. Many small teams do not. What they want is the outcome (a support inbox handled, leads followed up, a website kept current), and the workflow is only the means. For them the build-and-maintain cost from section 7 is the largest line on the real bill, larger than any platform fee, and it is the line no automation tool removes.

That is a different category of product, and it is where platforms such as O-mega sit: instead of a canvas of nodes, you describe the business or the function, and an AI builds and operates it (website, app, billing, content and the admin around them) through one conversation, billed in credits. It is not a replacement for a precise integration workflow with an audit trail, and a team that needs exact control of every step should stay with the tools above. But if the honest answer to "who will maintain these 40 Zaps?" is "nobody", it is worth comparing against a system that maintains itself. Our guide to AI agents for small businesses walks through which jobs suit which approach.

10. Outlook: Where Automation Billing Goes Next

The billing changes of the last fifteen months are not random repricing. They follow from the structural point in section 2: once the number of steps is decided by a model, the vendor has to choose who carries the variance, and the market is still settling on an answer. Reading the direction of travel matters if you are signing an annual contract, because the unit you budget in today may not be the unit you are billed in next year, as Zapier's Agents customers are discovering.

Four trends look durable enough to plan around. None of them is a prediction about a single vendor's next price change; each follows from forces that are already visible in the products and in the models underneath them.

10.1 Inference gets cheap, so the platform's share of the bill grows

The price of competent tool-calling inference keeps falling. Claude Haiku 5.5 lists at $0.10 per million input tokens, and OpenAI's GPT-6.1 Terra, released on October 26, was pitched as performing "on par with GPT-6 Astra at only one-tenth its price per token" - BleepingComputer. For the reference agent, a cheap model makes the token cost of a run well under a cent, at which point the platform fee becomes the largest line for any team that bothers to pick a small model.

That shifts pressure onto per-step platform pricing. A Zapier Premium multiplier of 5x makes sense when the bundled model is expensive; it looks steep next to a $0.004 token bill. Expect bundled-AI tiers to keep repricing toward cheaper default models, and expect more buyers to choose bring-your-own-key paths, which is why the 1x BYOK rate is the most important number on Zapier's AI page. Our live LLM API price table tracks which models make that switch worthwhile.

10.2 Automation platforms become tool layers behind other agents

The second trend is architectural. With MCP, the reasoning can live in Claude, ChatGPT, Copilot or your own code, and the automation platform becomes the layer of connectors and credentials that the agent calls. Zapier prices that role explicitly at two tasks per tool call, n8n exposes workflows as tools through its MCP Server Trigger, and Make runs an MCP server of its own. Pipedream went furthest: it is closing its workflow product to focus on being "the integration layer companies rely on to connect AI to the outside world" - Pipedream.

If that trend continues, the per-tool-call price becomes the unit that matters most, and the comparison between platforms becomes a comparison of connector breadth, credential security and cost per call. That favours Zapier's catalogue and n8n's flat economics over feature-rich canvases. Our overview of the 2026 MCP specification explains why stateless servers and long-running tasks make this layering cheaper to operate.

10.3 Enterprise software buys the orchestration layer

The third trend is consolidation. Celonis owns Make, Workday bought Pipedream in a deal announced in November 2025, and SAP took a stake in n8n in May 2026 at a $5.2 billion valuation, with plans to make n8n available inside its Joule Studio agent environment - Tech.eu. The logic is the same in each case: the system of record (process mining, HR, ERP) wants to own the layer where agents act on its data.

For buyers this cuts both ways. Integration with the parent's software gets deeper and better supported. Pricing, roadmap and even the product's survival follow the parent's strategy, as Pipedream Workflows users have just learned. When you choose a platform owned by or allied with a large vendor, read its roadmap as part of that vendor's, and keep the exit path from section 7.5 in mind.

10.4 Accountability becomes a product feature

The last trend is that someone has to answer for what an agent did, and that someone is increasingly the operator. Wells Fargo's agreement treating agent requests "as if they came directly from you" is one bank's version of a shift every regulated counterparty will make. The practical consequence is that per-tool approvals, audit logs of every tool call and narrow credentials move from nice-to-have to requirements, and the platforms that make them cheap will win regulated buyers.

Open weights add one more option at the edge. Mistral published the weights of the 1-trillion-parameter Mistral Large 4 in late October under a modified MIT licence - Mistral, which in principle lets a team run the whole stack (n8n plus a frontier-scale model) inside its own walls. In practice that takes serious GPU capacity, so for most teams the realistic path to data control remains self-hosted n8n with a hosted model under a data-processing agreement, or one of the smaller open models covered in our open-source LLM roundup.

11. The Decision Framework: Which Platform for Which Agent

The comparison reduces to a small number of questions about your agent and your team, not about the vendors. How many runs a month, and how many tool calls in each? Who will build and maintain it, and can that person run a server? Does the agent need a human to approve some actions? And how much would it cost you to move if the platform changed? Answer those four honestly and the choice usually makes itself, because each platform's meter rewards a specific shape of work.

The diagram turns those questions into a path. It is a starting point rather than a verdict: a team with strong reasons outside cost (an existing enterprise contract, a compliance requirement, a builder who already knows one tool) should weigh those first, since the switching and learning costs in section 7 are often larger than the price differences in section 6.

11.1 The short answers

For most teams putting AI agents into production in 2026, n8n Cloud is the best value: its per-run meter is the only one that does not grow with autonomy, it has had per-tool approval on every plan since January, and at our reference volumes it costs between $20 and $120 a month in platform fees. Self-hosted n8n is cheaper still at high volume and gives the most control over data and exit, but only if someone owns patching, a responsibility 2026's critical advisories made concrete. Make is the right answer for visual builders with modest volume and short agents, where its cheap credit beats n8n, provided they bring their own model key and avoid minute-by-minute polling.

Zapier remains the right answer for a different buyer: a non-technical team that needs its 9,000-app catalogue, values the simplest builder, and runs hundreds rather than tens of thousands of agent runs a month. For that team the premium buys real time savings. The mistake to avoid is accepting Zapier's defaults, because the Premium tier and monthly overage at 2.5x are where the price runs away. Teams that want outcomes rather than workflows should look past all three at agent-first products, or at autonomous-company platforms such as O-mega, covered in section 9.

11.2 Five moves that cut the bill on any platform

Whatever you choose, the same handful of decisions moves the real cost more than the choice of vendor. They come straight from the cost model in section 2: control the tokens, control the steps, and control the human time, because those are the three layers that grow.

None of these moves requires changing platform, which is the point: a team that switches vendors to save money but keeps a Premium default, a one-minute polling trigger and an approval on every lookup will often spend more on the new platform than it did on the old one. The levers below are listed in the order we would apply them to an existing agent, starting with the ones that take an afternoon and save the most.

1. Bring your own key and enable prompt caching; on Zapier, choose Advanced or your own key over the Premium default.
2. Route by difficulty: a small model for classification and lookups, a frontier model only for the step that needs it.
3. Cut wasted steps: webhooks instead of polling on Make, filters before AI steps, tool output trimmed before it reaches the model.
4. Gate writes, not reads: per-tool approval on actions with consequences, none on lookups.
5. Measure per run: tool calls, tokens, cost and reviewer minutes for every run, reviewed monthly.

The order matters. The first two moves act on inference, usually the largest variable cost, and can halve it in a day. The third acts on the platform meter, and matters most on Zapier and Make, where every step is billed. The fourth acts on human cost, which section 7 showed can exceed everything else combined. The fifth is what keeps the other four honest, because an agent's behaviour drifts as its inputs change, and a cost you do not measure per run is a cost you discover on the invoice. Our guide to seven ways to build an agent and what each costs covers the wider build-versus-buy picture if you are still deciding whether a platform is the right layer at all.

11.3 The verdict

The headline comparison of "n8n vs Zapier vs Make" is really a comparison of three billing philosophies. Zapier charges for actions, which is simple and predictable for deterministic work and expensive for agents, whose value lies in taking actions nobody scripted. Make charges for operations at a much lower unit price, and passes its own AI through at roughly the model's list price, which makes it cheap for short agents and less so for long ones. n8n charges for runs, which leaves you holding the token bill and the operational work but makes the platform side of an agent's cost almost flat.

Price the agent you actually have, not the one in the vendor's demo: count its tool calls, its tokens and its approvals, multiply by your volume, and add the people. Do that, and in most production cases the cheapest platform and the safest one turn out to be the same choice.

This guide reflects the pricing and features of n8n, Zapier, Make and the alternatives reviewed as of November 4, 2026. Make's prices vary by country and are shown for US visitors, and several features cited here (Make AI Agents, n8n's standalone agents) are in beta or preview. Automation pricing changes often, and Zapier has announced billing changes for December 1, 2026, so verify current details on each vendor's pricing page before committing.

What this article rests on

Every model figure above is read from its record, re-read from OpenRouter's model feed every 6 hours, so the figures on this page are current even where the prose around them is older. Prices are OpenRouter's listed price per million tokens: for a lab's closed models that is the lab's own list price, and for an open-weight model it is the price of the host OpenRouter routes to by default, which can differ from the lab's own API and from other hosts.

• GPT-6.1 Terra (OpenAI): $1 in and $5 out per million tokens, cached input $0.05, a 1.05M token window, listed 26 October 2026.
• Claude Haiku 5.5 (Anthropic): $0.10 in and $0.50 out per million tokens, cached input $0.01, a 1M token window, listed 7 October 2026.
• GPT-6 Luna (OpenAI): $0.10 in and $0.50 out per million tokens, cached input $0.01, a 1.05M token window, listed 22 September 2026.
• Claude Sonnet 5.5 (Anthropic): $2 in and $10 out per million tokens, cached input $0.10, a 1M token window, listed 28 September 2026.
• Claude Opus 5.5 (Anthropic): $4 in and $20 out per million tokens, cached input $0.20, a 1M token window, listed 22 September 2026.

In this topic: AI costs

• Muse Code Pricing 2026: Plans, Real Limits, Wonder Full
• Codex Pricing 2026: Plans, Limits, Resets and Credits
• Prompt Caching: Cut AI Agent Bills 45% in 2026
• Grok Bot Pricing 2026: Real Costs, Limits, Alternatives
• GPT-6 Astra Pricing: Real Cost per Agent Task 2026
• Gemini 3.8 Flash: $0.58/Task for Agents 2026
• Cheapest LLM APIs for AI Agents: 2026 Price Table
• Grok 4.6: Cheapest Frontier LLM for Agents 2026

What do AI models and agents cost, and how do you cut it? Every guide on it