Disclosure: some links on this page are affiliate links. If you sign up through one, we may earn a commission at no extra cost to you. We are not in Zapier’s or n8n’s affiliate programme, so nothing on this page earns us anything when you click through to either of them.
Every automation tool we cover now ships an AI feature. The feature is easy to find. What it costs you is not, because each company decided separately what to measure and where to put the meter.
We hold an account on all five. On October 25, 2026 we read their screens and then read what each company publishes about the charge. This page is about the shape of the bill, not about which assistant writes better prompts.
Quick Answer
Two of the five charge AI out of the allowance you already track. On Zapier an AI step costs 1 task on the Standard tier, 3 on Advanced and 5 on Premium, out of the same monthly task count that pays for your ordinary actions. On Make an agent run costs one credit per operation plus credits for the tokens it processes, out of the same credit balance every other module draws on. What happens once that allowance is gone is a separate comparison.
The other three put AI on a meter of its own. Pabbly keeps a separate AI counter beside the task allowance, Latenode reports Plug and Play tokens next to its CPU seconds, and n8n splits the charge twice over: one kind of credit for talking to its assistant, another for the model calls your workflow makes while it runs.
Only some of them tell you the price before you press run. Zapier prints the task cost next to each model tier and itemises a finished run in Zap history. Latenode itemises afterwards instead, giving each run its own columns for CPU seconds, Plug and Play tokens and operations. Make shows nothing while you are building, then prints the exact figure the moment the run ends: we ran one agent on the free plan and the module came back marked one operation, 1.05 credits.
Two Products Under One Word
Each of these companies sells two different things under the word AI, and they are not billed the same way. The first table is the assistant that writes a workflow for you. The second is the step that runs inside a finished workflow, every time it fires. Mixing them is how a comparison ends up telling you the wrong number.
The assistant that builds the workflow
| Tool | What it is called | Reachable without paying? | What it draws on |
|---|---|---|---|
| Make | Maia | A thirty-day trial from signup on the free plan | Paid plans for continued use |
| Pabbly Connect | Build with AI, which opens Pabbly AgenticAI | We opened the builder on the free plan; whether it is the AI Assistant named on the pricing page is not established | An AI counter separate from the task allowance |
| Latenode | Build with AI | No plan requirement printed | Plug and Play tokens at $1 each |
| Zapier | Copilot | Yes, with a daily message limit | No separate charge documented for the conversation |
| n8n | n8n Assistant | No permanently free Cloud plan | Assistant credits, which a workflow run never touches |
The AI step inside a running workflow
| Tool | Reachable without paying? | What the run draws on | Cost shown before you run? |
|---|---|---|---|
| Make | Yes — every plan, with Make’s own provider | The ordinary credit balance, plus token credits, plus a credit for each tool called | Nothing before the run; exact credits after it |
| Pabbly Connect | Not established — we opened the builder, not a run-time step | An AI counter separate from the task allowance | Balance only |
| Latenode | The free plan lists the AI node; using it bills tokens | Plug and Play tokens at $1 each, on top of CPU seconds | After the run — a column per figure |
| Zapier | Preview only, in an unpublished Zap | The ordinary task allowance, multiplied by model tier, with each tool call multiplied too | Yes — beside each tier, and itemised afterwards |
| n8n | No permanently free Cloud plan | Gateway credits, or nothing extra if the node uses your own credentials | Not on the workspace we hold |
Neither table has a cost column, on purpose. A task, a credit, a CPU second and a token are not the same quantity. Some companies publish a rate inside their own unit — Make converts tokens to credits, Latenode prices a token in dollars — but none of them publishes a conversion across the five. Putting them in one column would invite an arithmetic that nobody can do.
The second table is the one that matters for a workflow you intend to leave running. The first only tells you what it costs to get it built.
What This Comparison Covers
We hold an account on all five tools and read their screens on October 25, 2026, alongside what each company documents about AI charges. The subject here is the billing mechanism: which meter moves, whether you can reach the feature on a free plan, and whether the interface tells you the cost. What it costs to keep an ordinary two-step automation running, with no AI in it, is a separate question we answered by building the same workflow on all five: Google Sheets to Gmail, five ways. Where a figure comes from a vendor page rather than a screen we opened, the Sources section says so.
Builder and Run Step Are Two Different Charges
This is the distinction that makes the rest of the page readable, and it is the one most comparisons skip.
Every one of these products has two separate AI surfaces. One is an assistant that builds a workflow for you from a sentence: Zapier calls it Copilot, Make calls it Maia, n8n calls it the Assistant. The other is a step that sits inside a finished workflow and runs on every execution: AI by Zapier, the Make AI Agent node, Latenode’s AI node. Which of the assistants a free account can open at all, and where each company puts that limit, is in four AI builders and their free-plan gates.
They are billed on different meters, and n8n says so in the plainest terms of the five. Its documentation states that running a workflow doesn’t use Assistant credits, even if n8n Assistant built it — only conversations with the assistant do. The model calls a workflow makes while it runs come out of a second pool called Gateway credits.
The practical consequence: an assistant that builds your automation for free tells you nothing about what that automation will cost every time it fires. Those are two questions with two answers.
Make: On Every Plan, With the Rate Kept Elsewhere
Make’s help centre, read October 25, 2026, says AI Agents is available on all plans when you use Make’s own AI provider, and that connecting your own provider such as OpenAI or Anthropic requires a paid plan. The feature is in open beta, and Make notes that beta functionality and pricing may change.
That matches what the free account shows on the plan we use day to day, covered in our Make review. The AI Agent node drops onto the canvas with Make’s AI Provider already selected, and two model tiers are available: Small and Medium, both running gpt-5-nano. Larger tiers and the named OpenAI models are marked Paid.
The charge comes out of the same credit balance as everything else you run. Make’s help centre gives it as one credit per operation plus credits based on AI tokens when you use its own provider, and a tool the agent calls costs its own credit on top. Pick your own provider instead and the token half disappears, leaving the per-operation credit. A free Make plan carries 1,000 credits a month and allows two active scenarios.
The rate itself is published. Since August 25, 2026 Make has listed separate input and output rates per tier: one credit buys 18,080 input tokens or 2,260 output tokens on Small and Medium, and 3,616 input or 452 output on Large. Its own worked example is a data extraction step on Medium using 2,000 input and 100 output tokens, which it puts at about 0.16 credits under those rates.
So Make is among the easiest of the five to reach, and the number arrives after the fact rather than before it. We ran one to see. A single agent step on the Small tier, given a short message to summarise, came back with its header marked 1 operation · 1.05 credits · 1.5 KB — printed in the module’s own result panel, before we had left the editor. The organisation’s credit log carried the same 1.05 a moment later, and the balance on the dashboard moved by the same amount. Make does not break out how many tokens that bought, so the published per-tier rates are not something a run lets you check. That run called no tools. When we attached one and ran it again, the panel and the bill stopped agreeing — the readings are in what one agent run costs on Make’s free plan.
Make’s free plan is the one to open if you want to put an AI step on a canvas today without a card.
Pabbly Connect: A Counter of Its Own
Pabbly’s AI builder opens on the free plan — we reached it in October 2026 — and it draws on an AI credit balance that is not the task allowance.
Pabbly is unusually clear about the ordinary unit it sits beside, which is the part we went through in the Pabbly Connect review. Its task history screen states that any action performed in your workflow is considered a task, that triggers are not included in the task count, and that internal applications such as filters, routers and formatters are not counted either. It also notes that the task history only goes back fifteen days.
What one AI credit represents is a different matter. Pabbly does print the wallet. Its AI credits screen lists fifty credits as the all-time total, 7.02 of them spent by the single builder request we sent, and 42.98 left. The heading on that total is lifetime purchased, not a monthly figure, and the Units counter sitting beside it on the same screen is marked as remaining this month — so the two meters in the same product refill differently. More credits are sold at a hundred for five dollars. What Pabbly does not state is what one credit measures, so this page says only that the AI builder runs on a counter of its own, out of a balance that does not come back.
Pabbly Connect is the one to look at if you want the task meter itself to be legible, whatever the AI counter does.
Latenode: Line Items Per Run, and One Run That Never Appeared
Latenode’s execution history breaks a run into more columns than any other screen we opened: the time, the environment, the status, the duration, the CPU seconds, the Plug and Play tokens and the operations, each one separate. When a run is listed there, that is where its cost is.
The tokens in that column are not incidental. Latenode prices a Plug and Play token at one dollar and says it is billed in addition to workflow runtime, charged by actual provider usage rather than per request, and its AI nodes sit in that Plug and Play family. The free plan lists the AI Agent node, but neither page we read prints a free token allowance.
Its usage statistics page sorts consumption by scenario and labels it Scenario Execution. The CPU-second meter itself, and what a plain two-step run costs on it, is in our Latenode review. There is no separate line for polling or waiting.
There is a caveat we would want to know about if we were choosing on the strength of that reporting. On October 25, 2026 we added a row to the sheet a deployed Latenode scenario was watching. The email arrived about thirty seconds later, with the subject and body carrying values from that row, so the scenario ran. Filtering the history to production executions returned No executions match the current filters, and the CPU-second figures were unchanged fifteen minutes afterwards. We saw this once, and we do not know whether it was reporting lag or a class of run that is not counted.
Latenode’s free plan lists the AI node, and every run gets its own row of figures. Budget for the tokens separately: Latenode prices a Plug and Play token at a dollar and bills it on top of the runtime.
Zapier: One Allowance, With a Multiplier
Zapier folded AI into the unit it already had. Its help pages give the rate as a multiplier on a task: Standard counts 1, Advanced 3, Premium 5, and bringing your own model key counts 1. The plan table lists AI by Zapier on Professional and above, so none of the three tiers runs in a published Zap on Free. Standard has no tool support.
The part worth reading twice is how tool calls are counted. Zapier publishes the arithmetic:
Tasks used per run = (1 × model rate) + (number of tool calls × model rate)
Their own worked example is a Premium step that makes two tool calls: (1 × 5) + (2 × 5) = 15 tasks. The step is not five tasks with the tool calls thrown in. Each call is charged at the tier’s full rate.
The task itself is the unit we measured in the Zapier review. On a free account you can still put an AI step into an unpublished Zap and preview what it returns. The tier is fixed to Standard there, and the Zap cannot be published. That is what we saw when we added one: the editor flagged the draft as using a Pro feature.
One more thing is in motion. Zapier is moving its standalone Agents product into this step. Its migration page says each converted agent becomes a Zap with a native trigger and a single AI by Zapier step, that the pricing is the same task-based pricing as the rest of Zapier with no separate subscription, and that Zapier has not set a date to turn off Agents yet. If you are comparing write-ups from earlier in 2026, this is the thing most of them describe as a separate product.
n8n: Two Layers of Credits, and No Free Cloud Plan
n8n divides the AI charge more finely than the others. Assistant credits pay for conversations with its assistant, measured against the tokens the model processes for each message; a message you cut short or that errors still consumes what it used. Those credits reset at the start of each calendar month and do not roll over.
Gateway credits are the other half. They cover the model and service calls a workflow makes while it runs, and n8n’s documentation adds the line that matters most for anyone holding their own API key: nodes using your own credentials are not affected when the Gateway balance reaches zero.
A third name appears on the pricing page: the trial line reads 800 Assistant credits plus 100 AI credits, and the page never says what an AI credit is or lists one against a paid plan. We are leaving that where n8n left it.
Both named credit pools sit outside the execution quota that pays for running workflows, which is the meter we looked at in the n8n review. n8n’s pricing page lists 1,600 Assistant credits a month on Starter and up to 9,600 on Pro, with a trial allowance of 800. There is no permanently free n8n Cloud plan, so on this one the AI arrives with a trial or a subscription. What that absence looks like beside a tool that does have a free plan is in our Zapier vs n8n comparison.
We could not read any of this from a running instance. Our own workspace is paused after its trial, and the instance address returns a 404; what we could open was the Cloud console, which reports that Assistant credits are not available for that workspace yet and that the feature is still rolling out. The figures above are n8n’s own, not a balance we watched move.
Which One Should You Use?
If you want to try an AI step before paying anything, Make is the shortest route — the node runs on any plan with Make’s own provider, and the free plan carries a thousand credits a month. Work out the cost from the published token rates before you build, because the node will not show it to you.
If you already hold an API key from OpenAI or Anthropic, it is worth more on some of these than on others. On n8n a node using your own credentials does not draw Gateway credits at all. On Make it removes the token half and leaves one credit per operation. On Zapier it changes nothing about the Zapier charge: bringing your own key is still a task per step.
If you need to answer for the bill afterwards, Latenode and Zapier are the two that gave us a per-run breakdown. Latenode’s rows carry the CPU seconds, the Plug and Play tokens and the operations in separate columns; its free plan lists the AI node, though using one bills tokens at a dollar each on top of the runtime. Zapier prints the multiplier beside the tier before you build and breaks a run down by tool call, model tier and step afterwards, but you will need a paid plan to publish the Zap that uses it.
And if the AI step is the point of the automation rather than a garnish, read the unit before the price. A tool that charges per workflow run and a tool that charges per action behave differently once an AI step starts calling tools, because on Zapier each of those calls carries the tier’s multiplier.
For what it is worth about these links: we are affiliates for Make, Pabbly and Latenode. Zapier does not run an open programme and we are not in n8n’s, which is why the Zapier and n8n links here are plain. We have tried to keep that out of the ordering above — on the one measure where the five genuinely separate, which is whether the interface tells you the cost, the tools we earn from do not come first.
Everything on this page is an automation platform with AI inside it. The platforms where the product itself is the agent meter differently again, and we put five of those side by side in what AI agent platforms charge for.
FAQ
Does a failed AI step still cost me?
Zapier states that failed MCP tool calls do not count toward task usage. Its AI by Zapier pricing page does not make the same statement about AI steps, so we are not carrying that answer across from one product to the other.
Do all Zapier products charge one task per action?
No. Zapier publishes product-specific rates: a successful MCP tool call counts as 2 tasks, a successfully routed lead in Lead Router counts as 5, and a Code step uses 1 per run at standard runtime and more if you extend it.
If the assistant builds my workflow for free, is the workflow free to run?
No. These are separate meters. n8n puts it directly: running a workflow does not use Assistant credits even if the Assistant built it. The run draws on whatever the platform charges for execution, plus whatever its AI step draws.
Which of the five can I reach without a paid plan?
Make’s AI Agent node on any plan with Make’s own provider; Pabbly’s AI builder, which we opened on the free plan; and Latenode’s free plan, which includes the AI node. Zapier allows a preview in an unpublished Zap at the Standard tier. n8n Cloud has no permanently free plan. What the four with free plans did when we tried their built-in AI, and why there was nothing to try on n8n, is in which of these run AI on a free plan.
Can I compare the cost of an AI step across two of these tools?
Not directly. A task, a credit, a CPU second and a token measure different things. Make publishes a token-to-credit rate and Latenode prices a token at a dollar, but those are rates inside each company’s own unit, not a bridge between them. What you can compare is the shape: which allowance the charge comes out of, and whether a second one has to be budgeted alongside it.
Sources and Verification
Read on screen in accounts we hold, October 25, 2026: the Zapier editor, where an AI by Zapier step prints its tier and task cost and the draft was flagged as using a Pro feature; the Zapier app sidebar, showing Free Plan, Tasks 0 / 100 and Zaps Unlimited; the Make dashboard, showing a free subscription with 1,000 monthly credits and two active scenarios, and the AI Agent node with its model tiers and no cost figure; the Pabbly Connect task history, which prints the definition of a task and the fifteen-day window; the Pabbly AgenticAI plans and usage screen, reached from Build with AI in the Connect dashboard, showing available 42.98, lifetime purchased 50.00 and lifetime used 7.02 against Units marked 0 of 100 remaining this month; the Latenode execution history and usage statistics, including the run that delivered an email and did not appear under production executions; and the n8n Cloud console, reached after signing in, which shows our workspace paused and reports that Assistant credits are not available for it yet. We did not reach a running n8n instance.
Vendor documentation, read October 25, 2026:
- AI by Zapier model tier pricing — the 1x, 3x and 5x multipliers, the bring-your-own-key rate, the formula Tasks used per run = (1 × model rate) + (number of tool calls × model rate) with its fifteen-task example, and the free-plan preview in an unpublished Zap.
- How is task usage measured in Zapier? — one task per successful action, two per successful MCP tool call with failed calls not counted, five per routed lead in Lead Router, and one per Code run at standard runtime.
- Migrating from Agents to AI by Zapier — the move into a step in the editor, the same task-based pricing with no separate subscription, and Zapier has not set a date to turn off Agents yet.
- Zapier pricing — the free plan’s task allowance and what Copilot is available on.
- Make AI Agents app — available on all plans using Make’s AI Provider, with the option to use custom AI provider connections on paid plans, and the open beta notice that functionality and pricing may change. This page marks itself as describing a previous version, so the charges below come from the current pages.
- Credit usage for AI agents — one credit per operation plus token credits on Make’s provider for running an agent, a credit per operation for each tool called, and the custom-provider rows that drop the token component.
- Updated Make AI Provider token pricing — in force from August 25, 2026: 18,080 input or 2,260 output tokens per credit on Small and Medium, 3,616 input or 452 output on Large, and the vendor’s own worked example quoted above.
- Introduction to Maia — the builder’s plan requirements and the thirty-day trial for free plan users.
- n8n Assistant credits — credits measured on the tokens each message processes, running a workflow doesn’t use Assistant credits, even if n8n Assistant built it, the monthly reset and the absence of rollover.
- n8n Gateway credits — what they cover at run time, and that nodes using your own credentials are not affected when the balance reaches zero.
- n8n executions — the execution quota that sits apart from either credit pool. The absence of a permanently free Cloud plan is read from the pricing page above, which lists only a trial and paid tiers.
- n8n pricing — 1,600 Assistant credits a month on Starter, up to 9,600 on Pro, and 800 with the trial.
- Latenode pricing — the free plan line items including the AI Agent node, and the Plug and Play token at 1 token = $1, charged by actual provider usage, not just per request and billed in addition to workflow runtime.
- Latenode: what are credits — One PnP token costs $1.00, and that PnP nodes bill tokens on top of CPU seconds.
Figures described as on screen come from accounts we hold. Figures described as vendor documentation come from the company’s own pages on the date given. We did not run an AI step on any of the five for this page, so no per-run cost appears above.
Comments