Zapier's AI Step Pricing Is Eating No-Code Budgets — Here's How to Audit Yours

Zapier flipped the switch on June 15. AI steps stopped being simple task consumers and became something closer to pay-per-use API calls with a tiered multiplier baked in. Six weeks in, the real-world impact is clearer than anyone expected: teams are burning through tasks 3x to 5x faster on the same automations they were running in May.

The pricing change wasn't hidden. Zapier published a clear help article and gave notice. Standard tier: 1x base rate, no tool calls. Advanced tier (the default for new steps): 3x base rate plus additional task consumption for every tool invocation. Premium: 5x, with more sophisticated reasoning and tool support. Plus a per-step limit of 75 tasks before the Zap pauses and asks for manual approval.

The problem isn't transparency. It's that the default behaviours are expensive by design, and most teams haven't adjusted. If you're running AI-powered Zaps, this is the week to audit your configuration. Here's how.

How the new pricing actually works (in plain terms)

Every AI step now has a base cost determined by your model tier. If your plan allocates 750 tasks per month, a single Advanced-tier AI step that makes three tool calls consumes 3 (base rate at 3x) plus 3 (tool calls at 1x each). That's six tasks, minimum, for one Zap run. Run that Zap 125 times and your entire monthly allocation is gone. On a single automation.

Premium tier pushes the base rate to 5x. That same three-tool-call step now costs eight tasks. A Professional plan with 750 tasks handles 93 runs before you're out. Not 93 Zaps. 93 runs of one Zap.

Then there's the 75-task hard limit per step. If an AI step reaches 75 tasks in a single run — easy to do with iterative agent workflows that make multiple tool calls per decision cycle — the entire Zap pauses and requires manual approval to continue. For automations designed to run unattended, that's not a safety net. It's a silent failure mode. Your customer onboarding Zap pauses halfway through because the AI step got chatty with your CRM. Nobody notices until the customer emails support asking why their account still isn't set up.

The BYOK escape hatch and when to use it

Zapier offers a Bring Your Own Key option at 1x base rate. You connect your own OpenAI or Anthropic account, use their API directly, and Zapier charges you one task per run regardless of model tier. The AI provider bills you separately.

This is the right move for any team running more than a handful of AI steps per month. The arithmetic is straightforward: if your AI step volume is high enough that Zapier's 3x multiplier would push you to the next plan tier, BYOK almost certainly saves money, even after paying the model provider directly.

But BYOK has its own overhead. You're managing API keys. You're monitoring usage across two billing systems. You lose Zapier's unified billing and some of the simpler onboarding flows. For teams with one or two AI-powered Zaps, it's probably not worth the complexity. For teams running AI across a dozen workflows, it's the only sane option.

One thing worth checking: some models are cheaper through BYOK than through Zapier's Premium tier even before the multiplier. GPT-4o-mini via OpenAI's API costs a fraction of what Zapier's Premium tier charges in task-equivalent terms. If your AI steps are doing straightforward classification or extraction that doesn't need frontier reasoning, BYOK with a cheaper model beats every Zapier tier.

The audit framework: four questions to ask this week

If you're running AI steps on Zapier, here's what to check before the next billing cycle.

One: how many of your AI steps are on Advanced or Premium tier, and do they actually need to be? Advanced is the default for new steps because it enables tool calls. But if you're running a simple text extraction or classification that doesn't call external tools, Standard tier (1x) works fine and costs a third as much. Go through every AI step and ask: does this actually make tool calls? If not, switch to Standard.

Two: are your AI steps making more tool calls than necessary? Every tool invocation adds to the task count. If your step calls Google Calendar, then Gmail, then Slack, each of those is another task. Combine tool calls where possible. Where you can't, split multi-step AI workflows into separate Zaps so the expensive tool-heavy steps are isolated from the simpler ones.

Three: would BYOK save you money? Calculate your current AI step task consumption and multiply by the tier multiplier. Compare against the cost of calling the model provider directly plus the 1x Zapier task rate. The crossover point where BYOK wins is lower than most teams assume. I'd ballpark it at roughly 300 AI step runs per month — below that, the complexity isn't worth it. Above that, it almost always is.

Four: are you hitting the 75-task ceiling without realising it? Check your paused Zaps. If the AI step is silently pausing, those automations aren't running. You're not saving money. You're losing functionality and don't know it.

The architectural alternative: predictable AI pricing

There's a reason I keep coming back to platforms that handle AI pricing differently. Stacker's AI pricing is per-record operation, not per-model-call. You know what a workflow costs before it runs because the pricing model doesn't change based on which tier you selected or how many tool calls the agent decided to make. That predictability matters when you're building workflows that touch real business processes — invoicing, customer onboarding, compliance checks — where cost unpredictability isn't just annoying, it's actively dangerous for budget planning.

The broader point: per-call AI pricing, multiplied by tier, multiplied by tool invocations, creates costs that are functionally impossible to forecast. Any automation platform that ties AI pricing to model tiers is incentivising you to use simpler models than the task requires. That's not a good alignment of incentives between the platform and the builder.

Zapier's AI pricing change isn't unreasonable. The old model — treating a GPT-4 call the same as a text formatting step — wasn't sustainable either. But the implementation, with Advanced as the default tier and tool calls stacking multiplicatively, punishes teams that don't actively manage their configuration. If you haven't audited your AI steps since June 15, this week is the time.