The 2026 Automation Pricing Reset: Credits, Metered AI and What It Really Costs Now

The old rules everyone quietly relied on

For most of the last decade, pricing an automation stack was a countable exercise. The billing units were flat and boring, which is exactly what made them useful. Zapier charged you per task, where a task was any single action a Zap performed. Make charged per operation, where an operation was roughly one module firing. Power Automate billed per flow run when premium connectors were involved. You could look at how many times a workflow ran, multiply by how many steps it had, and arrive at a number that was close enough to plan around. Cost tracked volume, and volume was something a business could forecast.

That predictability was the foundation of an entire buying habit. Agencies quoted retainers against it, founders budgeted against it, and comparison articles ranked tools on it. The reason the 2026 landscape feels unsettled is not that any single platform got dramatically more expensive. It is that the thing everyone was counting stopped being the thing that drives the bill. AI broke the link between how often a workflow runs and how much it costs, and every major platform has now rewritten its pricing to account for that.

Make: from operations to credits

Make made the most explicit change. On 27 August 2025 it retired operations as its billing unit and replaced them with credits. On the surface this looked like a rename, and the company was careful to say that plans, prices and limits stayed the same, with unused operations converting to credits at a clean one-to-one ratio. For a team running classic no-code automations, the switch really was close to invisible: standard triggers, filters, HTTP modules and app connectors still consume one credit per execution, exactly as one operation did before.

The substance is in what credits let Make do that operations could not. A flat operation could only ever cost one unit, which made no sense once a single module might call a large language model or chew through a hundred-page PDF. Credits are variable by design. Language-model modules now charge based on token volume, so a longer prompt or a wordier response costs more. Document and media modules charge by file size. PDF and analysis modules charge per page. And running Make Code, the platform's JavaScript and Python step, costs two credits for every second of execution time. The billing unit did not just get a new name; it gained the ability to scale with the real cost of the work, which is precisely what AI made necessary.

Zapier: the AI-step multiplier

Zapier kept its familiar task model on the surface. The Professional plan still starts at 29.99 US dollars a month for 750 tasks, dropping to around 19.99 dollars a month on annual billing, and the bill still climbs with volume, landing somewhere in the region of 120 to 150 dollars a month once you reach 5,000 tasks. What changed is what counts as a task. From 15 June 2026, AI by Zapier steps are priced by model tier rather than as a flat single task. A standard model costs one task per run, the advanced tier, which is now the default, costs three tasks per run, and a premium reasoning tier costs five.

That multiplier is easy to underestimate. A workflow that once cost one task per run can now cost three or five if it routes through an AI step on the default settings, which means an AI-forward automation consumes your task allowance several times faster than the deterministic version of the same job. Zapier softened the blow elsewhere, making its Canvas planning workspace free on every plan and folding AI agents into existing plans with their own free activity tiers, so the platform is cheaper for planning and experimentation. But the meter that actually drains your allowance now runs faster whenever real AI work happens inside a live Zap.

n8n: execution billing and the open-source lever

n8n spent 2025 and 2026 turning its billing model into a marketing weapon, and the timing was deliberate. The company raised 180 million dollars at a 2.5 billion dollar valuation in October 2025, led by Accel with participation from Nvidia's venture arm, on the back of roughly 40 million dollars in annual recurring revenue. The pitch behind that raise is simple: n8n bills per workflow execution, not per step. A workflow can branch into twenty modules and still count as a single execution, whereas the same logic on a task-based platform might register as twenty separate billable units.

At low volume the difference is academic. At scale it is the whole ballgame. Published comparisons put the saving at more than 9,000 dollars a year at 50,000 monthly runs and more than 23,000 dollars a year at 100,000 runs against task-based pricing, which is why n8n is often described as ten to twenty times cheaper for complex, high-volume automations. On top of that, the self-hosted Community Edition is free and runs unlimited executions on a server that can cost as little as a few dollars a month, with a startup program at 400 dollars a month for small, early companies that want managed enterprise features. The catch is honest to state: self-hosting swaps a licence bill for operational work, and AI model calls still cost money on any platform because the tokens are billed by the model provider. We break that specific trade-off down in our guide to cloud versus self-hosted automation.

Power Automate: free until it isn't

Microsoft's model is the one that catches people out most often, because it looks free until a single step tips it into paid territory. Flows that use only standard connectors are effectively bundled into a Microsoft 365 subscription, which makes Power Automate genuinely cheap for a team already living inside Microsoft's ecosystem. The line is drawn at premium. The moment a flow touches a premium connector, uses desktop automation, or does anything beyond basic task automation, it needs a paid plan.

Those paid options come in two flavours. Power Automate Premium is 15 US dollars per user per month billed annually, which suits teams that want predictable per-seat costs. Pay-as-you-go bills premium cloud flow runs at roughly 0.60 dollars per run, which suits sporadic or spiky usage. Neither is expensive in isolation, but the licensing boundary is subtle: a business can build what feels like a free automation, add one premium connector to reach a database or a third-party app, and discover that the whole flow now requires a paid licence. The cost is real; it is just hidden behind a connector classification rather than a run counter.

Four platforms, four billing units

The clearest way to see the reset is to line up what each platform actually charges for, because the units are no longer comparable in the way a simple price-per-month table would suggest.

Read down the third column and the temptation is to rank by entry price. Read down the fourth and the real 2026 story appears: the deterministic core of each platform is priced roughly the way it always was, and the AI layer bolted onto it is where the volatility lives. That is the pattern every one of these companies converged on independently, and it is the single most useful thing to understand about your bill this year.

Why the bill got unpredictable

The uncomfortable truth behind all four models is that a language-model step does not have a fixed cost. Its price depends on how many tokens the prompt and the response contain, and that varies with every input. A customer who writes three sentences costs a fraction of a customer who pastes an entire email thread. A summary of a short document is cheap; a summary of a long one is not. When the cost of a step is a function of its content rather than its existence, the old habit of estimating a bill by counting runs simply stops working.

Every platform's response to this is a version of the same move: meter the AI. Make meters it in credits tied to tokens and file size. Zapier meters it with task multipliers per model tier. Power Automate meters it through AI Builder credits and Copilot licensing. n8n pushes the meter entirely onto the model provider, where the tokens are billed directly. The mechanism differs but the consequence is identical. Two runs of the same automation can now cost different amounts, and a busy month with heavy inputs can cost far more than a quiet one, even at the same run count.

This is also why so many teams report that their automation return on investment came in below expectation in 2026: the model they budgeted against assumed flat per-run costs that no longer exist. We dug into that gap between promise and outcome in why automation ROI is lower than expected, and metered AI is one of the biggest new reasons the numbers drift.

Did automation get more expensive, or cheaper?

Both answers are true at once, which is why the debate feels muddled. On a unit basis, AI automation got dramatically cheaper. The per-token price of capable models fell hard as open-source options matured and platforms competed, and some estimates put the drop in AI automation pricing at around 35 percent between 2024 and 2026. The raw ingredients of an AI workflow cost meaningfully less than they did two years ago.

Yet plenty of teams saw their total bills rise, because falling unit costs collided with exploding consumption. When an AI step becomes a two-click affair, people add far more of them, and each one carries a multiplier or a credit draw the old deterministic step did not. Add-on AI agent tiers priced at 150 to 200 dollars a month on top of a base plan can push an AI-heavy stack above the task-based bill it was supposed to replace. Cheaper parts, used far more often, can still add up to a bigger invoice. If you want the fuller picture of what a modern AI workflow costs end to end, our breakdown of the total cost of ownership of an AI agent walks through the line items the sticker price hides.

How to budget in the new model

You cannot go back to counting runs, but you can make consumption behave. The teams that kept their 2026 bills predictable did a handful of unglamorous things consistently.

1. Split deterministic steps from AI steps. Track them separately. The deterministic part of your workflow is stable and easy to forecast; the AI part is the volatile line item that deserves the attention.
2. Right-size the model tier. The default is rarely the cheapest. On Zapier that means checking whether a standard model at one task will do the job instead of the advanced default at three. Reserve premium reasoning for the steps that genuinely need it.
3. Trim the tokens. Because cost scales with input and output length, shorter prompts and capped response lengths translate directly into a smaller bill. Do not feed an entire document to a model when a relevant extract will do.
4. Set hard limits on metered usage. Any pay-as-you-go meter, such as Power Automate's per-run billing, should have alerts and a ceiling so a runaway loop cannot quietly spend real money overnight.
5. Model against peak, not average. Overage rates punish the busy month. Size your plan to a realistic high-volume month so a spike does not tip you into expensive per-unit territory.
6. Re-measure before you migrate. Pricing changes are a reason to check your actual usage against your platform's shape, not an automatic reason to switch. Rebuilding and re-testing a stack usually costs more than a monthly saving unless your volume is genuinely large.

The through-line is that the lever moved. It used to sit on how many times things ran, and you controlled it by controlling volume. Now it sits on how much your AI steps consume, and you control it by controlling model choice and input size. For a wider view of where each tool fits once the pricing dust settles, our comparison of n8n versus Make versus Zapier maps the platforms onto the kinds of work they suit best.

What the reset means going forward

The 2025 to 2026 repricing is not a temporary wobble that will settle back into flat per-run charges. It is the industry catching up to a permanent change in what automations do. As long as workflows contain steps whose cost is a function of their content, billing will stay variable, and the platforms that win will be the ones that make that variability visible and controllable rather than the ones that hide it. Make's move to credits, Zapier's tiered AI multipliers, Power Automate's connector line and n8n's execution model are four different bets on how to price the same underlying reality.

For the businesses buying automation, the practical lesson is smaller and more durable than any single price change. Stop treating your automation bill as a fixed subscription and start treating it as a consumption budget with a stable base and a variable AI layer. Watch the variable layer the way you would watch a cloud compute bill, because that is what it has become. Do that, and the 2026 reset stops being a source of nasty surprises and becomes just another line you manage on purpose.

FAQ

What changed with Make's pricing in 2025?

On 27 August 2025 Make replaced operations with credits. Standard modules still cost one credit per run and unused operations converted one-to-one, so simple workflows were largely unaffected. The real change is that AI, file and code modules now charge variable credits based on tokens, file size, page count or seconds of execution, so an AI-heavy scenario can dominate the bill.

How much does Zapier cost in 2026?

The Professional plan starts at 29.99 US dollars a month for 750 tasks, or around 19.99 dollars monthly on annual billing, rising toward 120 to 150 dollars at 5,000 tasks. Since 15 June 2026, AI steps are metered by model tier at one, three or five tasks per run, so AI workflows drain the allowance faster than deterministic ones.

Why did automation pricing become harder to predict?

Because AI steps do not have a fixed cost. Their price depends on how many tokens each input and output contains, which varies every run. Platforms met that by metering AI in credits, task multipliers or tokens, so the same automation can cost different amounts on different days, and budgeting shifted from counting runs to modelling consumption.

Is n8n really cheaper than Zapier and Make?

For complex, high-volume workflows, usually yes, because n8n bills per full execution rather than per step, saving a reported 9,000 dollars a year at 50,000 runs and 23,000 dollars at 100,000 runs versus task pricing. But AI tokens are billed by the model provider regardless of platform, and self-hosting trades a lower licence cost for running your own server.

How does Power Automate bill for automations?

Flows using only standard connectors are effectively free inside Microsoft 365. Touch a premium connector or desktop automation and you need a paid plan: Premium is 15 US dollars per user per month billed annually, and pay-as-you-go premium runs cost about 0.60 dollars each. A single premium step can pull an otherwise free flow into paid territory.

Did automation get more expensive or cheaper overall?

Both. Unit prices for AI automation fell by roughly 35 percent between 2024 and 2026, but teams added far more AI steps, and add-on agent tiers at 150 to 200 dollars a month pushed many total bills up. Cheaper ingredients, used much more often, still produced bigger invoices for a lot of businesses.

Should I switch platforms because of the pricing changes?

Not on price alone. Rebuilding, re-testing and retraining usually cost more than a monthly saving unless your volume is genuinely large. Treat the changes as a prompt to re-measure your real usage against your platform's billing shape, and switch only when the numbers clearly justify it.

How can I keep my automation bill predictable in 2026?

Track AI steps separately, use the cheapest model tier that works, trim prompt and response length to control tokens, set alerts and hard limits on pay-as-you-go meters, and size your plan against a realistic peak month rather than an average one so a spike does not trigger overage rates.