Zapier AI models have multiplied, and the variety quietly changed how automations should be built. Open the app’s AI step these days and you get a wall of names: GPT this, Claude that, Kimi, GLM, Muse Spark. Picking wrong doesn’t break your Zap, but it can triple your task cost or ship sloppy output to every customer on your list. Here’s the short version of which model fits which job.
First, understand the two routes
Zapier gives you two ways to wire AI into a workflow, and the choice shapes everything after it.
Route one: AI by Zapier. One built-in step that packages dozens of models from OpenAI, Anthropic, Google, and more. No separate account, no API tokens, no surprise invoices. You pay in Zapier tasks, which makes budgeting almost boring. For beginners, this is the route.
Route two: direct integrations. You connect a provider’s own account, often with your API key, and get a wider lineup as a reward: legacy GPT versions, DeepSeek, and others that never made the built-in menu. Billing happens by token through the provider. More control, more homework.
There’s also a third door that isn’t a model at all: Zapier MCP installs Zapier inside MCP-compatible AI chats, so you can kick off automations across 9,000+ tools by typing in your chat window. Worth knowing, not worth starting with.
The task tiers, translated
Zapier sorts its AI models into three tiers based on the underlying model’s cost. This is the pricing table that matters, so here it is cleaned up:
| Tier | Cost per run | Examples | Best for |
|---|---|---|---|
| Standard | 1 task | GPT-5 mini, GPT-4o mini, GPT-6 Luna, Claude 5.5 Haiku | Summaries, labels, replies, most text jobs |
| Advanced | 3 tasks | GPT-4.1, Claude 4.5 Haiku | Longer context, trickier formatting |
| Premium | 5 tasks | GPT-5, GPT-4o, Claude 5.5 Opus, Claude 4.8 Opus | Nuanced writing, complex reasoning |
Here’s the math beginners miss. A Zap that runs 500 times a month costs 500 tasks on a Standard model and 2,500 on a Premium one, for work a mini model usually handles fine. Start Standard. Upgrade only when output quality actually drops.
Which model for which job
Now the part you came for. Match the work, not the hype:
Summarizing, labeling, routing, and replying. A Standard model does this in its sleep. GPT-5 mini or Claude 5.5 Haiku will classify support tickets and draft replies all day at 1 task a pop.
Writing anything customer-facing. This is where I’d pay for Premium. Claude Opus models write with a voice, and a 5-task charge on a lead-gen email beats a flat-sounding Standard draft you have to fix by hand. If you automate 20 emails a month, the premium bill is pocket change.
Multi-step agent work. Kimi, from Moonshot AI, was purpose-built for tool use and long chains of actions. Zapier’s own AutomationBench leaderboard currently ranks Kimi K3 #36 overall, at $0.43 per task. For agentic Zaps that hop between apps, that mix of capability and price is hard to beat.
High-volume grunt work. Z.ai’s GLM models are the budget kings: GLM 5.2 costs $0.22 per task, one of the cheapest options on the board. It ranks #90 on AutomationBench, so it’s not your reasoning engine of choice. For straightforward jobs at scale, though, cheap and fine beats brilliant and broke.
Technical pipelines. DeepSeek is the open-weight option with strong coding and reasoning chops. Developers wiring technical workflows on a budget tend to land here.
Already on AWS or Azure? Then you have enterprise routes: Amazon Bedrock runs models inside the security boundary your IT team already vetted, and Azure OpenAI keeps data in your private Azure tenant, which is why regulated industries pick it. Neither applies to a solo business, and that’s fine.
The picky details worth knowing
Two features round out the picture.
Bring your own key (BYOK). If the built-in menu lacks a model you want, BYOK lets you plug a provider key into AI by Zapier and use it inside the same step. You still pay Zapier tasks for the run, plus the provider’s token charges. It’s the middle ground between the two routes.
Model churn is real. The Zapier AI models catalog shifts as providers ship. The mini models of six months ago weren’t the mini models of today. Once a quarter, open your Zaps that use AI steps and sanity-check whether a newer Standard model can replace a pricier pick. Ten minutes, real savings.
One caution from the field: resist automating anything you haven’t done manually at least once. The model choice matters less than knowing what good output looks like. Our walkthrough of five Zapier Sidekick automations shows that principle in practice, and the guide to automating ChatGPT with Zapier covers the no-code basics if you’re brand new to this.
Three mistakes people make picking models
Since we’re being honest about the field, here are the errors I see most:
Defaulting to Premium for everything. People grab the strongest Zapier AI models “just to be safe,” then wonder why 4,000 tasks evaporated in a week. Safety costs money here. Tier down, check the output, tier up only if it’s genuinely bad.
Judging a model from one run. Models are streaky. A single weird reply doesn’t condemn a Standard model, and one great paragraph doesn’t justify Premium. Run the same prompt three or four times on each candidate before you commit a busy Zap to it.
Forgetting the human check. An AI step inside a Zap runs unsupervised, forever, on live data. Whatever model you pick, add a filter or a review step before anything posts publicly or emails a customer. Automation scales your judgment, including the bad days.
The clear winner for beginners
If you want one answer: use the AI by Zapier step with a Standard-tier model, almost always GPT-5 mini or Claude 5.5 Haiku. It’s the cheapest capable setup, it requires zero provider accounts, and it survives model churn because Zapier maintains the menu. Every Zapier AI models guide will argue about the frontier picks, but beginners win by defaulting simple. Move a specific Zap to Premium when its output disappoints, and look at Kimi or GLM only when volume makes cost sting.
Your next move
Open one Zap you already run and check which model the AI step uses. If it’s Premium and the job is summarizing or labeling, downgrade it and watch your task usage drop. Then re-check the Zapier AI models lineup next quarter, because the leaderboard never sits still: Zapier keeps its own updated guide worth bookmarking. And if you want a model that watches social feeds, our Grok automation walkthrough covers that angle.