Zapier Just Changed AI Pricing Too — 48 Hours After Anthropic. Here's What No-Code Automators Need to Know

June 15, 2026 was not a coincidence. On the same day Anthropic split Claude's billing into subscription and metered tiers, Zapier rolled out an identical structural change: AI steps are no longer flat-rate. The model you pick sets the price. The more tools your step calls, the faster your task allowance burns. And every AI Zap you built before that Sunday? Frozen. It still runs, but it can't use tools or switch models.

Two platforms. Same day. Same underlying logic: the free AI ride is over, and the meter started running everywhere at once.

This isn't just a Zapier story. It's part of a triple squeeze that landed inside 72 hours: Anthropic on June 15, Zapier on June 15, and Google killing its free Gemini CLI entirely on June 18. We covered the Gemini shutdown yesterday. This piece is about what Zapier changed, what it means for your automations, and the pattern connecting all three.

What Zapier actually changed

Before June 15, an AI by Zapier step consumed one task per run. Whether it called GPT-4.1 nano or Claude Opus, whether it used zero tools or five, the meter ticked once. That's gone.

AI steps now use model-based pricing with four tiers:

| Tier | Task multiplier | Tool support |

|---|---|---|

| Standard | 1x | No |

| Advanced | 3x (default) | Yes |

| Premium | 5x | Yes |

| Bring Your Own Key | 1x | Yes |

New steps default to Advanced, meaning a simple AI step that used to cost you one task now costs three before it does anything useful. The task formula is straightforward, and it gets expensive quickly:

Tasks per run = (1 × model rate) + (number of tool calls × model rate)

A Premium-tier step that calls two tools (say, a CRM lookup and a Slack message) hits 15 tasks in a single run: one for the step itself at 5x, plus two tool calls at 5x each. The same workflow under the old system was one task. That's a 15x increase.

Zapier added a 75-task runaway limit per run. If a step hits that ceiling, the Zap pauses and waits for your approval. It's a guardrail, not a cap. The bill still comes.

What's frozen and what you can do about it

Legacy AI steps, anything created before June 15, still run at the old 1-task rate. Zapier isn't upgrading them automatically. But they're permanently locked: no tools, no model selection, no tier switching. They're museum pieces that still work.

If you want tools, knowledge sources, or the ability to pick models, you must build a new AI step. That new step joins the metered world.

There is one escape hatch: Bring Your Own Key. Connect your own OpenAI, Anthropic, Google, Azure OpenAI, or Amazon Bedrock account, and you get 1x task consumption with full tool support. You pay Zapier for the task and your AI provider for the tokens. For tool-heavy workflows that call multiple apps per run, BYOK is the cheapest option by a wide margin. The trade-off is that you're now managing two bills instead of one.

The Anthropic parallel, same day, same pattern

On June 15, Anthropic stopped subsidising agent usage. Interactive Claude Code and chat stayed on your subscription. But automated usage (CI pipelines, headless Claude Code, third-party agents) moved to a separate credit pool: $20/month for Pro, $100 for Max 5x, $200 for Max 20x. Exceed the credit and you pay API rates per token. Zed Industries estimated subscriptions were subsidising agent use by 15 to 30 times.

Compare the two side by side:

| | Anthropic (June 15) | Zapier (June 15) |

|---|---|---|

| What changed | Agent/automated usage split from subscription | AI steps priced by model tier |

| Old cost | Flat subscription covered everything | 1 task per AI step run |

| New cost | $20–$200 monthly credit, then per-token | 1x–5x tasks per run, plus per-tool-call |

| Who's hit hardest | CI/CD pipelines, headless agents, bots | Tool-heavy Zaps, Premium-tier users |

| Escape hatch | Use interactive Claude Code (stays on sub) | Bring Your Own Key (1x tasks) |

Same shape. Different products. The playbook is identical: identify heavy automated users, separate them from the flat-rate pool, meter them individually.

What this costs in practice

Take a real-world Zap: inbound email arrives, AI step categorises it, AI step drafts a reply, Slack step sends a notification. Under the old system, the two AI steps cost two tasks. Under the new system, with both AI steps on Advanced tier and no tools, they cost six tasks. Three times the burn rate with identical functionality.

Add tools and the maths gets worse. A Premium AI step that calls a Notion lookup, a Google Sheets append, and a Slack message burns 20 tasks per run (1 base at 5x + 3 tool calls at 5x each = 20). Run that 50 times a day and you've consumed 1,000 tasks on AI steps alone before any of your other Zaps get a look in. On the Professional plan (750 tasks/month included), you'd exhaust your allocation in under a day.

BYOK drops that same workflow to 4 tasks (1 base at 1x + 3 tool calls at 1x). The difference between Premium and BYOK for a tool-heavy Zap is the difference between a sustainable automation and an invoice you don't want to open.

The third blow: Google Gemini CLI

Google killed its free, open-source Gemini CLI on June 18. 100,000 GitHub stars, 6,000 community pull requests, a year of contributor labour, and the replacement is a closed-source Go binary called Antigravity CLI. No feature parity. No grace period. Just errors where `gemini -p` used to work, as we detailed in yesterday's piece.

Three companies. Three decisions. 72 hours. The timing isn't coordination, it's convergence. Every AI platform is arriving at the same conclusion simultaneously: free infrastructure was a land grab, the land is grabbed, now monetise.

What every Zapier user should do this week

An audit checklist that takes about an afternoon:

1. Find every AI by Zapier step in your account. Zapier doesn't surface this in one dashboard, so check Zap by Zap. Note which ones use tools, which ones run at high volume, and which ones you forgot existed.

2. Decide on each legacy step. If it's a simple prompt (classification, summarisation, extraction) and you're happy with the default model, leave it. It costs one task and it'll keep costing one task. Don't touch what isn't broken.

3. For tool-heavy Zaps, switch to BYOK immediately. The maths is unambiguous. Premium with three tool calls: 20 tasks. BYOK with three tool calls: 4 tasks. The token cost from your AI provider is almost certainly less than the 500% task premium Zapier charges on Premium tier.

4. Downgrade where possible. Not every AI step needs Claude Opus. If your step is doing something a GPT-4.1 mini or Gemini 2.5 Flash Lite can handle, drop to Standard tier and save 67% to 80% on tasks versus Advanced.

5. Set usage alerts. If you're on Team or Enterprise, configure task usage notifications. The 75-task per-run guardrail prevents a single runaway Zap from eating your whole month, but it won't stop death by a thousand small overruns.

The AI tax is arriving everywhere, all at once

There's a phrase that keeps coming up when you look at these three changes together: the AI tax. It's not a single company's pricing decision. It's the market-wide pattern of free and cheap AI infrastructure being withdrawn, re-bundled, and metered.

For no-code builders, the lesson isn't "avoid AI." It's that the platforms where you embed AI now carry a pricing volatility you didn't sign up for six months ago. The Zap you built in May at 1 task per AI step might cost 15 tasks today if you rebuild it on Premium with tools. The Claude pipeline you scripted in April might now be drawing from a $20 credit pool you've already burned.

Structured platforms with managed AI, Bubble, Webflow, Stacker, the Zapier platform itself when you use it without the new AI step metering, absorb this volatility. Their engineering teams handle model migrations and pricing negotiations. When you're wired directly to provider pricing (Zapier's new AI tiers, Anthropic's credit pools, Google's CLI whims), you absorb it yourself.

That's not a reason to abandon AI automation. It's a reason to be deliberate about where you sit in the stack. For high-volume, tool-heavy workflows, BYOK is the cheapest and most durable option right now. For simple classification and extraction, leave legacy steps alone. For everything else, check the maths before your next invoice does it for you.