Can a ChatGPT or Claude subscription run your automations? What vendors say

By Rafal Fila, ShadowGenRead Guide

Abstract

A guide to the three ways a workflow can pay for an LLM step: a subscription login, an API key, or platform credits. It quotes what vendor terms and billing pages say, prices one synthetic step in each platform's unit, and sorts steps by credential. It gives no ruling on what is allowed.

Limits within limits on a subscription

Subscription plans meter use in nested windows, and OpenCode Go allows 20% of its monthly limit per five hours and 50% per week. OpenCode Go

Claude limits use per five hours as well as per week, Anthropic so an automation that runs all day will meet these windows, whereas an API key bills by the token.

ClaudeOpenCode GoWeekly resetOne turn is one weekFive-hour windowMonthly capOpenCode GoWeekly resetOne turn is one weekFive-hour windowMonthly cap
This figure in words
  • Two wire coils rise from ground ellipses. Height is elapsed time, and one turn of a coil is one week.
  • Claude's coil carries a tick every five hours and a ring at each turn end. Its paid plans limit use per five-hour session and per week, and the weekly limit is shared with Claude Code. Anthropic
  • OpenCode Go's coil carries the same ticks and rings and stops at one month. It allows 20% of the monthly limit per five hours, 50% per week and 100% per month. OpenCode Go
  • Claude's coil runs on past the month, dashed, because its page names no monthly cap.
  • An automation that runs all day meets these windows, whereas an API key bills by the token.
  • Kimi Code keeps a rolling five-hour window and removed the weekly quota for new members. Kimi Z.ai and MiniMax are in the window matrix below.

Windows named on each plan's own page, read 16 Nov 2026. AnthropicOpenCode Go Height is time, drawn to one month, and the shape shows how the windows nest rather than how much each allows.

Key figures

Zapier default tier Premium, 5x checked
A new AI step on a paid plan
Premium, 5x
Starts on Premium, at 5 tasks a run with no tool calls.
Standard
1x tasks
Advanced
3x tasks
Your own key
1x, Advanced or Premium models
Source help.zapier.com, AI by Zapier model tier pricing. Updated 13 Nov 2026. Read . Zapier model tiers
Zapier cap per run 75 tasks checked
Task limit for one AI step run
75 tasks
The default is 75 tasks, and you can raise it to 500.
Default
75 tasks
Highest
500 tasks
Source help.zapier.com, AI by Zapier model tier pricing. Updated 13 Nov 2026. Read . Zapier model tiers
Power Automate premium 10 credits checked
Copilot Credits per 1K tokens, premium
10 credits
Basic is 0.1 and standard is 1.5, on the same tokens.
Basic
0.1 per 1K tokens
Standard
1.5 per 1K tokens
Pay-as-you-go
$0.01 a credit
Source learn.microsoft.com, AI Builder message management and licensing. Read . Microsoft rate table
Claude paid plans 5-hour limit checked
Session limit on Claude paid plans
5-hour limit
A session limit resets every five hours, beside a weekly limit.
Weekly limit
shared across Claude and Claude Code
Source support.claude.com, usage and length limits. Read . Claude usage limits

Three credentials can pay for one step

Your workflow can pay for an LLM step in three ways. You can use a person's subscription login, an API key you hold, or credits the platform sells you, and each of them carries its own rules about who may use it and from which tool.

This page reports what the vendors' own pages say and gives no verdict on what is allowed, so you should read each source before relying on it, particularly when a job runs with nobody watching.

How four platforms connect to a model

The four platforms count an AI step in four different units.

Zapier counts a task, Zapier Make a credit per operation, Make n8n an execution n8n and Power Automate a Copilot Credit. Microsoft Learn

Zapier, Make, n8n and Power Automate each offer a key route as well as a credits route, whilst none of the four pages describes a subscription login as a connection method.

Model connection routes, from vendor documentation read 16 Nov 2026. Sources: n8nMakeZapierMicrosoft Learn
PlatformYour own API keyBundled creditsUnit it counts
n8nEvery plan and editionGateway credits, Cloud Starter and Pro onlyOne execution per whole run
MakePaid plans; you pay the provider for tokensMake's AI Provider, all plans1 operation is 1 credit for third-party AI apps
Zapier1x tasks, Advanced or Premium models onlyTiers of 1x, 3x and 5x tasksTask, scaled by tier
Power AutomateBring-your-own-model through FoundryCopilot Credits per 1K tokensCopilot Credit

What vendors say about subscription logins

Two of the eight vendor pages, GitHub's and OpenCode Go's, say nothing about unattended use.

OpenAI and Google both point automation at an API key instead. OpenAIGemini CLI

What each of the eight vendor pages says
What each vendor's page says about using a subscription outside its own app, read 16 Nov 2026. Sources: AnthropicOpenAIGemini CLIGitHubKimiZ.aiMiniMaxOpenCode Go
VendorWhat the page says
AnthropicDoes not permit routing requests through Free, Pro, Max or Ultra plan credentials on behalf of users. Your own key in your own environment is covered if usage bills to the key owner. Third-party tools may draw on usage credits, not plan limits. Signed in with a plan, the Agent SDK and claude -p still draw from plan limits.
OpenAI"The right way to authenticate automation is with an API key". Plan usage in other apps needs a separate permission per app.
GooglePiggybacking on Gemini CLI's OAuth from third-party tools is "a direct violation"; the supported route is an API key.
GitHubOfficially supports a Copilot subscription with OpenCode. Unattended use: Not stated.
KimiThird-party tools and API keys count towards the same quota.
Z.aiQuota may not be used for "directly invoking model APIs from your own applications, bots, websites, SaaS products". Calls outside the plan are not available.
MiniMaxPlan is "designed for individual, interactive developer use"; pay-as-you-go is recommended for production.
OpenCode GoDesigned for OpenCode and other coding agents that send similar requests. Unattended use: Not stated.

The wording splits in two, as the big model vendors mostly point automation at an API key, whereas the coding-plan vendors say what their plan is built for.

Anthropic's two pages point different ways. The legal page bars third-party developers from routing requests through plan credentials for their users. Anthropic The support page says the Agent SDK and claude -p still draw from plan limits when you are signed in. Anthropic A 15 Jun 2026 update paused the previously announced changes to Agent SDK usage. Anthropic

Both pages are quoted here, with no verdict.

Since 7 Oct 2026, Max, Ultra and Team plans include monthly API credits, Anthropic which cover the Agent SDK and claude -p when you run them with an API key from your linked Console organisation.

Rate windows and scheduled runs

All five subscription plans read here meter usage in a 5-hour window.

Four also meter a week, though Kimi removed that for new members, Kimi and only OpenCode Go adds a monthly cap. OpenCode Go

Every plan meters a 5-hour window. Hover a plan for its page's wording.Claude paid plans5-hour session limit, weekly limit shared with Claude CodeZ.ai Coding Planquota refreshes on a 5-hour and a 7-day cycleKimi Coderolling 5-hour window; weekly quota removed for new membersOpenCode Go20% of the monthly limit per 5 hours, 50% per weekMiniMax M Plan5-hour and weekly windows for text models5-hourWeeklyMonthlyClaudeClaude paid planssession limitshared with Claude CodeZ.aiCoding Plan5-hour cycle7-day cycleKimiKimi Coderolling windowremoved for new membersOpenCode GoShare of the monthly limit20%50%100%MiniMaxM Plantext modelstext modelsSubscription plan readUsage window the plan metersDot: the page names that window. Ring: the page says it was removed. Line: the page read names no such window.Every plan meters a 5-hour window.Hover a plan for its page's wording.Claude paid plans5-hour session, shared weekly limitZ.ai Coding Plan5-hour and 7-day refresh cyclesKimi CodeRolling 5 hours; no weekly for new membersOpenCode Go5 hours 20%, week 50%, month 100%MiniMax M Plan5-hour and weekly windows for text models5-hourWeeklyMonthlyClaudePaid planssessionsharedZ.aiCoding Plancycle7-dayKimiKimi CoderollingremovedOpenCode GoShare of month20%50%100%MiniMaxM PlantexttextPlan readUsage window it metersDot: page names the window. Ring: removed.Line: no such window on the page read.
This figure in words
  • Claude's paid plans have a five-hour session limit and a weekly limit shared across Claude and Claude Code. Anthropic
  • Z.ai's Coding Plan refreshes its quota on a 5-hour cycle and a 7-day cycle. Z.ai
  • Kimi Code keeps a rolling 5-hour window and has removed the weekly quota for new members. Kimi
  • OpenCode Go allows 20% of the monthly limit per 5 hours, 50% per week and 100% per month. OpenCode Go
  • MiniMax's M Plan applies 5-hour and weekly windows to text models. MiniMax

Usage windows named on each plan's own page, read 16 Nov 2026. No page read says what a scheduled job does when it hits one.

As of 16 Nov 2026, a rolling window suits a person who stops to think, and a scheduled job doesn't stop. Claude's paid plans have a five-hour session limit and a weekly limit shared across Claude and Claude Code, Anthropic whilst Z.ai runs a 5-hour cycle and a 7-day cycle. Z.ai

Kimi keeps a rolling 5-hour window and has removed the weekly quota for new members. Kimi OpenCode Go caps a model at 20% of its monthly limit per 5 hours and 50% per week, OpenCode Go and MiniMax applies 5-hour and weekly windows to text models. MiniMax

No page read says what happens to a scheduled job that hits a window, so ask the vendor before you schedule one. OpenAI's CI guide adds that one auth file should serve one serialized job stream. OpenAI Which matters if several scheduled jobs run at once.

One decision step in each unit

On the same tokens, 1,000 runs cost $2 on Power Automate's basic tier and $200 on premium.

Premium charges 100 times the credits of basic Microsoft Learn at $0.01 a credit, Microsoft Learn with standard sitting at $30 in between.

Power Automate, 1,000 runs of one AI step. Hover a tier for its credits.AI tool tier and its rateCost, creditsBasic tier0.2 credits a run, 200 for 1,000 runs, $2Standard tier3 credits a run, 3,000 for 1,000 runs, $30Premium tier20 credits a run, 20,000 for 1,000 runs, $200$1$10$100$1,000Basic0.1 credit per 1K tokens$2200 creditsStandard1.5 credits per 1K tokens$303,000 creditsPremium10 credits per 1K tokens$20020,000 creditsCost of 1,000 runs at 2,000 tokens each, US dollars, log scalePower Automate, 1,000 runs of one AI step.Hover a tier for its credits.Tier, rate per 1K tokensCost, creditsBasic tier0.2 credits a run, 200 for 1,000 runs, $2Standard tier3 credits a run, 3,000 for 1,000 runs, $30Premium tier20 credits a run, 20,000 for 1,000 runs, $200$1$10$100$1,000Basic0.1 credit per 1K tokens$2200 creditsStandard1.5 credits per 1K tokens$303,000 creditsPremium10 credits per 1K tokens$20020,000 credits1,000 runs at 2,000 tokens,US dollars, log scale
This figure in words
  • The basic tier charges 0.1 Copilot Credit per 1K tokens, so a run uses 0.2 and 1,000 runs use 200 credits, or $2. Microsoft Learn
  • Standard charges 1.5 credits per 1K tokens, so a run uses 3 and 1,000 runs use 3,000 credits, or $30. Microsoft Learn
  • Premium charges 10 credits per 1K tokens, so a run uses 20 and 1,000 runs use 20,000 credits, or $200. Microsoft Learn
  • Microsoft bills the tokens on all three tiers.

A synthetic AI step of 2,000 tokens, run 1,000 times, with every run succeeding. Dollars are at Microsoft's pay-as-you-go $0.01 a Copilot Credit, on a log scale where each gridline is ten times the last.

Cite as Automation Atlas, LLM subscriptions for automation, rates read 16 Nov 2026.

Power Automate charges 100 times as many credits for premium as for basic on the same tokens. Microsoft Learn At Microsoft's pay-as-you-go rate of $0.01 a credit, Microsoft Learn 1,000 runs cost $2 on basic, $30 on standard and $200 on premium. That spread is where an AI step gets expensive, and a headline plan price won't show it.

On your own key, a Zapier AI step costs 1 task, not 5 on the default tier, as of 16 Nov 2026. Zapier

Tasks per run are the model rate times one plus the tool calls, so with no tool calls, Premium's 5x default costs 5 tasks, whilst your own key's 1x costs 1.

Platform and routeUnits per run, one square eachFor 1,000 runsZapierStandard tier1 task1,000 taskstokens: ZapierZapierAdvanced tier3 tasks3,000 taskstokens: ZapierZapierPremium (the default)5 tasks5,000 taskstokens: ZapierZapieryour own key1 task1,000 taskstokens: you, to the providerMakeyour own key1 credit1,000 creditstokens: you, to the providern8nyour own key1 execution1,000 executionstokens: you, to the providerPlatform and routeFor 1,000 runsZapierStandard tier1 task1,000 taskstokens: ZapierZapierAdvanced tier3 tasks3,000 taskstokens: ZapierZapierPremium (the default)5 tasks5,000 taskstokens: ZapierZapieryour own key1 task1,000 taskstokens: you, to the providerMakeyour own key1 credit1,000 creditstokens: you, to the providern8nyour own key1 execution1,000 executionstokens: you, to the provider
This figure in words
  • On Zapier's own tiers a run costs 1 task on Standard, 3 on Advanced and 5 on Premium, the default, and Zapier bills tokens. Zapier
  • On your own key, Zapier charges 1 task a run, or 1,000 tasks for 1,000 runs, and you pay the provider for the tokens. Zapier
  • Make on your own key uses 1 credit a run, or 1,000 credits, and you pay the provider for the tokens. Make
  • n8n on your own key counts 1 execution a run, or 1,000 executions, and again you pay the provider for the tokens. n8n

Billing units for one LLM decision step, from rates read 16 Nov 2026. The scenario is synthetic: 2,000 tokens per run, 1,000 runs a month, no tool calls, every run succeeds. Triggers and other steps add to each total. One square is one unit, and Power Automate's three tiers are in the chart above. Sources: Zapier, Zapier Make, Make n8n, n8n Power Automate. Microsoft Learn Make's own AI Provider and n8n Gateway credits convert tokens by model, so they are not priced here.

On Zapier, each tool call the step makes adds the model rate again, so a Premium step with two tool calls uses 15 tasks. Zapier No vendor page gives a cost per completed task, so this table counts units per run.

Match each step to a credential

Different kinds of step call for different credentials, at least on the pages read, though this page makes no claim about which model does any of them better.

Only one of three step kinds points to a subscription, and that is code generation with a person present.

Bulk steps point to the lowest tier or your own key, Zapier whereas judgement steps point to a key you hold. Anthropic

One AI stepBulk orchestrationRouting, tagging, extracting fieldsCode generationDrafting or editing codeJudgement and decisionApprove, reject, escalateLowest platform tier,or your own keyA personpresent?A subscription insidethe vendor's own toolAn API keyin a pipelineYour own metered key,billed to youyesnoBulk orchestrationRouting, tagging,extracting fieldsLowest platform tier,or your own keyCode generationDrafting orediting codeA personpresent?A subscription insidethe vendor's own toolAn API keyin a pipelineyesnoJudgement and decisionApprove, reject,escalateYour own metered key,billed to you
This figure in words
  • Bulk orchestration, such as routing, tagging and extracting fields, points to the lowest platform tier ZapierMicrosoft Learn or your own key. Zapier
  • Code generation, such as drafting or editing code, with a person present points to a subscription inside the vendor's own tool. AnthropicMiniMax
  • Code generation in a pipeline points to an API key. OpenAI
  • Judgement and decision steps, such as approve, reject or escalate, point to your own metered key, billed to you. AnthropicOpenAIGemini CLI

Credentials the pages point to, by kind of workflow step. Sources: ZapierMicrosoft LearnAnthropicMiniMaxOpenAIGemini CLI The green route is the only one that ends at a subscription.

Bulk steps run often and cost little each, so the multiplier matters more than the model. Zapier's Standard tier is 1x, and Power Automate's basic rate is 0.1 credit per 1K tokens. ZapierMicrosoft Learn

Code generation is where subscriptions are built to be used: Anthropic describes its login as designed for ordinary use of its own apps, Anthropic and MiniMax calls its plan interactive. MiniMax Move that step into a pipeline, though, and OpenAI's guide points back to a key. OpenAI

Judgement steps decide who gets paid or told, so you want the billing and the limits to be yours, which a key you hold gives you. Anthropic That matters most when the output moves money.

What to do when credits are bundled

Zapier caps an AI step at 75 tasks a run by default, and you can raise it to 500.

The default tier matters too, because a new step on a paid plan starts on Premium. Zapier

Check the default tier as well as the price before you build, because as of 16 Nov 2026, a new Zapier AI step on a paid plan starts on Premium at 5x. Zapier

Set a cap per run, since Zapier's default is 75 tasks and you can raise it to 500. Zapier From 1 Dec 2026, AI by Zapier through Zapier MCP or the SDK uses its own price. Zapier

Know what stops at zero. n8n Gateway nodes stop until you add credit, while nodes on your own credentials carry on. n8n Power Automate agent flows, on the other hand, always consume Copilot Credits. Microsoft Learn

Bring your own model where you can, because Power Automate then charges no extra credits for processing and the token bill moves to Foundry. Microsoft Learn

How this differs from the coding plans guide

Atlas's AI coding plans guide compares what Claude Code, Codex, Copilot and Cursor plans cost and meter for a developer, whereas this page covers workflows with nobody there. For each platform's counting rules in dollars, see the cost by workload guide.

Questions

Can a Claude subscription run an unattended workflow?

Anthropic's legal page does not permit routing requests through Free, Pro, Max or Ultra plan credentials on behalf of users. Anthropic Its support page says the Agent SDK and claude -p still draw from plan limits when signed in. Anthropic Both are quoted here, with no verdict.

How many Zapier tasks does an AI step use?

Tasks per run are the model rate times one plus the tool calls. Zapier With no tool calls, Premium uses 5, Advanced 3 and Standard 1, whilst your own key runs at 1x. Zapier

What does an AI step cost in Power Automate?

Basic is 0.1, standard 1.5 and premium 10 Copilot Credits per 1K tokens. Microsoft Learn At $0.01 a credit, Microsoft Learn 1,000 runs of 2,000 tokens cost $2, $30 or $200.

Can I use my own API key on n8n or Make?

n8n allows your own AI provider key on every plan and edition. n8n On Make's paid plans you pay Make credits for operations and your provider for tokens. Make

What happens when a scheduled job hits a rate window?

No page read says, so ask the vendor before you schedule one. Every plan read meters a 5-hour window. AnthropicZ.aiKimiOpenCode GoMiniMax

Sources

Model and plan vendors

  1. Anthropic, Log in to your Claude accountPage undated. Read 16 Nov 2026.
  2. Anthropic, Use the Claude Agent SDK with your Claude planDated 7 Oct 2026. Read 16 Nov 2026.
  3. Anthropic, How do usage and length limits work?Page undated. Read 16 Nov 2026.
  4. OpenAI, CI/CD authenticationPage undated. Read 16 Nov 2026.
  5. OpenAI, Sign in with ChatGPTPage undated. Read 16 Nov 2026.
  6. Google, Gemini CLI FAQPage undated. Read 16 Nov 2026.
  7. GitHub, GitHub Copilot now supports OpenCodePage undated. Read 16 Nov 2026.
  8. Kimi, Kimi Code membershipPage undated. Read 16 Nov 2026.
  9. Z.ai, Coding Plan FAQPage undated. Read 16 Nov 2026.
  10. Z.ai, Subscription termsPage undated. Read 16 Nov 2026.
  11. MiniMax, M Plan FAQPage undated. Read 16 Nov 2026.
  12. OpenCode, OpenCode GoPage undated. Read 16 Nov 2026.

Automation platforms

  1. Zapier, AI by Zapier model tier pricingUpdated 13 Nov 2026. Read 16 Nov 2026.
  2. Make, CreditsPage undated. Read 16 Nov 2026.
  3. n8n, Gateway creditsPage undated. Read 16 Nov 2026.
  4. n8n, Can I use my own AI API key on n8n Cloud?Page undated. Read 16 Nov 2026.
  5. n8n, PricingPage undated. Read 16 Nov 2026.
  6. Microsoft, AI Builder message managementPage undated. Read 16 Nov 2026.
  7. Microsoft, Administer AI Builder licensingDated 8 Sep 2026. Read 16 Nov 2026.

Each chip in the text opens its source here. Only Zapier, Anthropic's Agent SDK page and Microsoft's licensing page printed a date.