AI Workflow Index

AI assistants can now build workflows. What should run them behind the scenes, and what will it cost?

Helping enterprise and mid-⁠market leaders decide where AI belongs, what to build next, and how governance should evolve

Get the PDF

Q3 2026 · The running cost of model-⁠led and deterministic AI orchestration

FOREWORD

Foreword

AI assistants can now build workflows, but something still has to run them. That is an architectural choice: a model that keeps checking for work (model-⁠led orchestration), or software that acts only when there is work to be done (deterministic AI orchestration).

Imagine you’re at home waiting for a courier to deliver an envelope containing instructions for your next job. You could get up periodically, turn on the lights, and check if the courier is there. Or, you could install a doorbell and stay on the couch until it rings. Either way, you read the instructions and get to work when the courier arrives.

In deterministic AI orchestration, you’re using “doorbells”. The workflow starts when something happens: a customer submits a form, a record changes, or a scheduled time arrives. Software handles the sequence of steps, calling AI when reasoning is needed.

In virtually all the configurations we tested under model-⁠led orchestration, you’re “getting off the couch periodically to check”. A model activates on a schedule, reads its instructions and context, and decides what to do and which tools to call. If there’s nothing there, it has still spent tokens finding that out.

On our central estimate, model-⁠led orchestration cost 2.1× as much overall. For workflows that complete fewer than 30 runs a month, it cost 8.3× as much.

Here is why. Most business workflows spend their days waiting and step in only when something needs attention. In this study, 61% of workflows completed fewer than 30 runs a month and accounted for about half the projected model-⁠led bill. Most of that group’s cost came from what we call sentinels: workflows whose checks usually find nothing to do, such as watching for a policy renewal.

The first AI Workflow Index examined where AI belongs inside a workflow. This edition measured the running costs of the two orchestration architectures. We rebuilt a smaller set of workflows as working automations on OpenAI and Anthropic’s business products and measured what they cost to run. We combined those measurements with observed workflow activity and estimated checking schedules to price all 2,944 workflows at published usage-⁠based rates.

AI can increasingly build either kind of workflow.

It can also handle the reasoning when something needs attention. For most workflows in this study, letting conventional software run the workflow cost less.

That points to a practical division of labor for enterprise leaders: use AI to build the workflow and handle reasoning. Let conventional software run it.

These findings reflect today’s pricing and product capabilities. We expect the two architectures to converge. Model-⁠led systems will increasingly gain ways to respond to events by building “doorbells”, while deterministic AI orchestration platforms already let users build and deploy workflows using AI.

As these capabilities mature and token prices change, the economics will shift. Reliability, governance, and the cost of building and maintaining workflows already matter; this edition does not measure them. Section 3 sets out how to evaluate your approach; the appendix shows how to price your own workflows.

SUMMARY

Model-⁠led orchestration cost more than twice as much

Six in ten AI workflows run infrequently, and that is where model-⁠led orchestration cost eight times as much; frequent workflows were close to even.

MODEL-⁠LED MONTHLY RUNNING COST RELATIVE TO DETERMINISTIC AI ORCHESTRATION

OVERALL2.1×2,944 workflows
30+ COMPLETED RUNS A MONTH1.2×1,153 workflows (39%)
UNDER 30 COMPLETED RUNS A MONTH8.3×1,791 workflows (61%)

Why it costs more

In the configurations tested, most model-⁠led workflows paid to check even when there wasn’t something to do. And the less often a workflow had something to do, the more of its bill went to empty checks. That is where the money went: workflows completing fewer than 30 runs a month made up about half the projected model-⁠led bill. Among workflows with billable tasks, deterministic AI orchestration cost less for about 4 in 5. The 2.1× is the central estimate; the checking-⁠frequency range is 1.3× to 3.7×.

What we compared

We analyzed a month of activity from 2,944 AI workflows and modeled their running costs at the published high-⁠volume task rate. We rebuilt a set of them on OpenAI and Anthropic’s business products, measured what they cost to run, and used those measurements with observed activity and estimated checking schedules to model costs across the full population. We compared only configurations available through built-⁠in features and connectors.

Deterministic AI orchestration: software follows predefined steps and rules, calling AI where needed. Model-⁠led orchestration: a model decides which actions or tools to use as the workflow runs. Multiples are model-⁠led monthly running cost ÷ deterministic AI orchestration for the same workflows, averaged equally across OpenAI and Anthropic; Zapier tasks are priced at $0.003, the monthly-⁠billed rate of the highest published Team tier, plus estimated provider charges for third-⁠party AI steps. Boundaries and assumptions are in the appendix.

TAKEAWAYS FOR LEADERS

  1. 01

    Don’t pay AI to look for work

    Make conventional software the default for starting workflows and running their steps, with AI handling the steps that need judgment.

  2. 02

    Evaluate the full cost of your approach

    Estimate running costs using a representative mix of workflows. Include engineering, maintenance, reliability, and governance in your organization’s build-⁠or-⁠buy decision. Revisit it as prices and capabilities change. The appendix provides a six-⁠step method for estimating running costs.

CONTENTS

SCOPE

Three drivers of AI workflow running cost

LAST EDITION

What each step does

Using AI selectively cost 71% less than routing every step through a model.

THIS EDITION

What runs it

Deterministic AI orchestration versus model-⁠led orchestration of the same AI workflows.

FUTURE EDITION

Which model is used

Not covered in this edition.

ABOUT THE DATA

Same workflows, same steps, compared on two architectures

This study compares what it costs to run the same AI workflows two ways: deterministic AI orchestration, which we priced on Zapier, and model-⁠led orchestration, which we priced on ChatGPT Workspace Agents and Claude scheduled tasks, averaged equally. It covers recurring running costs at enterprise scale. It compares configurations a nontechnical builder can create using built-⁠in product features and connectors, including native schedules and event triggers.

A workflow in this report is an automation which contains a trigger and a sequence of steps (at least one of them using AI) that ran in Q3 on Zapier.

We started with 2,944 AI workflows from leading Zapier accounts, each with a month of real activity: how often it ran and what it did. We rebuilt a smaller set as working automations on both model-⁠led products and measured what they cost to check for work, decide whether to act, and do it. We then applied those measured costs to all 2,944, using each workflow’s own activity and an estimated checking schedule.

The comparison preserves each workflow’s intended tasks and rules. The resulting cost differences reflect both how the tested implementations operate and how the products charge for them.

Prices are public list prices. Zapier tasks are priced at $0.003, the monthly-⁠billed rate of the highest published Team tier, plus estimated provider charges for third-⁠party AI steps. Model-⁠led usage is priced at the published rates for the tested products. Seats, build effort and maintenance are excluded.

One assumption came from a separate test: about 1,500 workflows had their AI steps re-⁠run on OpenAI and Anthropic models to see how often the smaller, cheaper model does the job. The appendix explains the result and how it is used.

These are modeled running-⁠cost estimates, not account bills. The study does not assess the value of flat-⁠rate AI subscriptions.

01

What runs the workflow

What runs the workflow?

Whichever AI assistant your teams build in, the question is what runs the workflow behind the scenes once it exists. Both architectures follow instructions a builder wrote. In one, software runs each step and calls a model where judgment is needed; in the other, a model reads the instructions and carries out each step itself. In this study, deterministic workflows start through software triggers; most model-⁠led workflows start a model session on a schedule to check for work.

Only model-⁠led orchestration pays to check for work

EXAMPLEA priority customer files a support ticket

Check for work

Has any work arrived?

Decide whether to act

Does it meet the criteria?

Do the work

Execute the AI step and make updates.

Deterministic

Check for work

Software trigger or poll

No model cost

A ticket is created; the help desk tells the workflow.

Decide whether to act

Software applies rules

No model cost

A filter checks the customer’s plan and the ticket’s priority.

Do the work

AI step + updates

Billed when run

A model drafts the reply; the workflow posts it and updates the record.

Model-⁠led

Check for work

Model wakes on schedule

Billed each check

Every hour a model session opens the queue to see whether a ticket has arrived. Most hours it has not.

Decide whether to act

Model applies criteria

Included in the check

The model reads the ticket and the criteria and decides whether this one qualifies.

Do the work

Same AI step + updates

Billed when run

The same model drafts the reply, posts it, and updates the record.

Both approaches can use AI to perform the same work. Deterministic orchestration uses software to find and filter incoming work. Model-⁠led orchestration typically starts a model session on a schedule to do this, consuming tokens even when no work has arrived.

AI can build either kind of workflow. The distinction is how the workflow operates after deployment. A noted exception is OpenAI’s Slack channel mode, described on the next page. The support-⁠ticket example is illustrative: the two approaches diverge in the first two steps and converge in the third.

In our tests, the model checks for work, even when there’s nothing to do.

Only the model could check a connected app for new work; the code it wrote could not. So every scheduled check used the model and was billed, even when it found nothing to do.

WHAT WE FOUND

Only the model can look inside your apps. Code the agent writes can’t check them on its own, so every check runs through the model and is billed. Our testing found the same inside the agent product.

SOURCE · OPENAI · MCP SERVERS AND CONNECTORS (API DOCUMENTATION)
“When the model decides to call an MCP tool, the API will make a request to the remote MCP server to call the tool and put its output into the model’s context.”

developers.openai.com, MCP servers guide, read November 2, 2026

WHAT WE FOUND

The sandbox where Claude runs code never holds your app credentials, so code it writes can’t check your apps. In our tests, only the model could, through a separate service, and every check was billed.

SOURCE · ANTHROPIC · COWORK ARCHITECTURE OVERVIEW
“Connector authorization tokens never enter the sandbox; connector calls are made on the server side.”

support.claude.com, article 14479288, read November 2, 2026

One exception in our tests: OpenAI’s Workspace Agents can start when a message arrives in a connected Slack channel. We priced each eligible Slack workflow at the lower of that trigger’s measured cost and the cost of scheduled checks. Workspace Agents also support API-⁠triggered runs, but connecting other systems to that API requires a separate integration, outside this study’s scope. See Boundaries of the finding.

02

What it costs, and why

Checking for work drives most of the model-⁠led bill

Model-⁠led orchestration cost 2.1× as much

as deterministic AI orchestration, on the same 2,944 workflows

Deterministic
1.0×
Model-⁠led average
2.1×

Averaged across OpenAI and Anthropic, model-⁠led orchestration cost 2.1× what deterministic AI orchestration costs to run the same workflows.

The cost is primarily in waking up to check for work. Checking frequency changes the estimate: halving or doubling the assigned checking frequency moves the combined result from 1.3× to 3.7×.

Deterministic AI orchestration: software follows predefined steps and rules, calling AI where needed. Model-⁠led orchestration: a model decides which actions or tools to use as the workflow runs. Model-⁠led average: ChatGPT Workspace Agents and Claude scheduled tasks, weighted equally. Deterministic: Zapier tasks at $0.003, the published high-⁠volume rate, plus estimated provider charges for third-⁠party AI steps. Bills vary by plan, and the multiple moves the opposite way to the task rate, almost in proportion. 2,944 workflows, including 113 that already use agent steps. The 2.1× is the central estimate. Builder-⁠stated schedules are held fixed in the checking-⁠frequency range; model-⁠choice assumptions move the result between 2.0× and 3.2× (see Sensitivities).

An estimated 8% of model-⁠led spending goes to doing the work

Checking for work and starting runs account for the other 92%.

Estimated allocation of each model-⁠led dollar · average of OpenAI and Anthropic

  • 60¢Checks that find nothing
  • 32¢Start-⁠up on checks that find work
  • 8¢Work

In the deterministic configuration priced here, checking and start-⁠up carry no separate usage charge. Billable tasks and estimated third-⁠party AI charges together come to about 48 cents for every $1.00 of modeled model-⁠led spending.

The allocation covers all 2,944 workflows and combines observed workflow activity with modeled checking costs: each workflow’s monthly cost is split by how many times it checked for work and how many of those checks found work, with a check that finds work priced at 1.2× an empty one; for Slack-⁠triggered workflows priced on OpenAI’s channel trigger, each message the workflow ignored counts as a check that found nothing. The shares are estimates, not vendor billing line items. Deterministic AI orchestration: Zapier at $0.003 a task plus estimated provider charges on third-⁠party AI steps. The overall 2.1× and this allocation cover the same 2,944 workflows (see Overall result and Methods).

Model-⁠led orchestration cost 8.3× as much for infrequent workflows.

Workflows completing fewer than 30 runs a month made up 61% of the sample; the multiple was 1.2× for more frequent workflows.

MODEL-⁠LED AVERAGE, MULTIPLE OF DETERMINISTIC

Frequent · 30+ completed runs a month
1.2×
Infrequent · under 30 completed runs a month
8.3×

The vertical line marks 1.0×, where model-⁠led and deterministic cost the same.

Frequent: 30 or more completed runs in the month (1,153 workflows). Infrequent: fewer than 30 (1,791 workflows). The cost of infrequent workflows depends strongly on how often the model checks between completed runs: a workflow checked hourly that finds work a few times a month pays for hundreds of empty checks.

Model-⁠led average: ChatGPT Workspace Agents and Claude scheduled tasks weighted equally. Zapier at $0.003 a task plus estimated provider charges for third-⁠party AI steps.

Model-⁠led cost more for 93% of infrequent workflows

Each curve ranks one group’s workflows from cheapest to most expensive on model-⁠led; above the dashed line, model-⁠led cost more.

Frequent 30+ completed runs a month · 1.2× overallInfrequent Under 30 completed runs a month · 8.3× overall
Model-⁠led cost / deterministic costLogarithmic scale
10,000×1,000×100×10×1×0.1×
Above 1×Model-⁠led is
more expensive
1× = Equal cost
Below 1×Model-⁠led is
less expensive
← Cheapest on model-⁠ledMost expensive on model-⁠led →

Dots mark where each curve crosses 1×: 42% of frequent workflows and 7% of infrequent ones cost less on model-⁠led.

Source: 2,645 workflows with a Zapier charge. Model-⁠led = average of OpenAI and Anthropic modeled costs; deterministic = Zapier tasks at $0.003 plus estimated provider charges on third-⁠party AI steps (113 agent-⁠step workflows: tasks only).

Model-⁠led cost more in three of four AI roles

Clerk workflows were roughly at parity overall; unusually high-⁠volume workflows influenced that result.

ROLEWORKFLOWSFREQUENTINFREQUENTALL
Communicatorwrites for people1,0054.0×32×8.1×
Clerkextracts and updates records*1,0160.62×3.9×1.0×
Analystdecides what happens next1223.2×3.5×3.3×
Coordinatorcreates tasks for teams1926.4×4.4×5.2×

One primary role per workflow. 609 workflows with no clear role are not shown. *Clerk results are sensitive to a concentration of unusually high-⁠volume workflows. Frequent Clerk workflows were the one group where model-⁠led cost less.

At 2,000 workflow scale, the difference is about $230K a year

HYPOTHETICAL ENTERPRISE: MONTHLY RUNNING COST FOR 2,000 WORKFLOWS USING THIS STUDY’S WORKFLOW MIX AND AVERAGE RUNNING COSTS

Deterministic AI orchestration · about $17,100 a month

Frequent $14,800
Infrequent $2,300

Model-⁠led orchestration · about $36,000 a month

Frequent $17,400Infrequent $18,600
Frequent · 780 workflows (39%)Infrequent · 1,220 workflows (61%), lighter shade

Over a year: about $205,000 deterministic against $432,000 model-⁠led. About 86% of the difference comes from the 1,220 infrequent workflows.

Average monthly cost per workflow from this study: frequent $19.00 deterministic and $22.34 model-⁠led; infrequent $1.84 and $15.24. Model-⁠led is the OpenAI and Anthropic average; deterministic is Zapier at $0.003 a task plus the estimated provider charges on third-⁠party AI steps. Bar lengths are in proportion. This scenario scales the observed portfolio averages, which are sensitive to unusually high-⁠volume workflows. The example scales this study’s mix and averages; an enterprise with a different workload or pricing would see a different figure.

03

Applying the finding

Putting software in charge of the workflow

The cost findings support a default for enterprise AI workflows: conventional software runs the workflow, and AI handles the steps that need judgment.

For leaders, the decision is how to provide that capability across the business, with consistent governance, access controls, and oversight. Will you build it yourself, or standardize on a platform? Compare running costs alongside reliability and ongoing ownership.

What your orchestration needs to support

A consistent division of work across your workflows, managed under shared governance and access controls.

STAGECONVENTIONAL SOFTWAREAIWHAT TO REQUIRE
StartingDetect events or check for new work without calling a model.Judge whether incoming information needs attention when rules alone are insufficient.Event triggers and scheduled checks that meet response-⁠time needs without unnecessary model calls.
DecidingApply predefined filters, routing rules, and thresholds.Score, classify, and make judgments where rules are insufficient.Clear limits on model decisions, with review or escalation when needed.
DoingExecute actions: write records, send messages, and create tasks.Draft content, extract information, and determine what an action should contain.Controlled access to connected systems, records of actions, and failure handling.

AI can help build workflows and handle bounded agent steps within them. The orchestration still needs maintained connections, retries, logs, permissions, and an owner, whether your team builds it or buys a platform.

This framework draws on the cost findings in this edition and the division of work described in the first Index. Governance, reliability, and maintenance are decision criteria, not measured outcomes of this study.

Build it yourself, or standardize on a platform?

Evaluate the needs of your workflow portfolio and the controls your organization requires.

REQUIREMENTBUILD INTERNALLY WHEN…STANDARDIZE ON A PLATFORM WHEN…
Systems and connectionsCritical systems or integration needs are not adequately supported by available platforms.A platform supports the systems and connections your organization needs.
Scale and response timeAvailable platforms cannot meet your operating requirements, and your team can build and sustain an alternative.A platform meets your expected volume, response-⁠time, and running-⁠cost requirements.
ReliabilityYou need custom failure handling and can operate the retries, monitoring, logs, and alerts yourself.A platform provides the reliability features and operational visibility you require.
Governance and accessYour controls require capabilities available platforms cannot provide, and your team can implement and maintain them.A platform provides consistent permissions, approvals, and auditability across your workflows.
Ongoing ownershipYou have a funded team responsible for maintaining the orchestration infrastructure and its connections.You want a vendor to maintain the orchestration platform and connectors, with your team owning the workflows.

Whether you build or buy, your organization remains responsible for workflow logic, access decisions, and responding to failures.

The study measured running costs, not engineering or maintenance costs. Include both in the organization-⁠wide build-⁠or-⁠buy decision.

04

Implications for leaders

Don’t pay AI to look for work

Set a default for how AI workflows run across the business, then evaluate the full cost of supporting it.

  1. 01

    Make deterministic AI orchestration the default

    Standardize on an approach that uses conventional software to detect incoming work and run the workflow, calling AI where judgment is needed. Use event triggers to avoid unnecessary model calls, under shared governance and access controls. AI can still build workflows and handle bounded agent steps within them.

  2. 02

    Validate the economics across your workflow portfolio

    Use the six-⁠step guide in the appendix on a representative mix of workflows, including frequent workflows and sentinels whose checks often find nothing. Use those estimates to inform your organization’s build-⁠or-⁠buy decision. Include engineering, maintenance, reliability, and governance alongside running cost, and revisit the decision as prices and capabilities change.

05

Acknowledgments

AUTHORS · ZAPIER RESEARCH LAB
  • Andre Vanier

    Head of Zapier Research

  • Gary Zukowski

    Senior AI Automation Strategist

  • Chris Ladouceur

    Senior Product Designer, Growth & AI

  • Anja Simic

    Head of Marketing, Enterprise

DESIGN, PRODUCTION, AND EDITING
  • Ryan McCready

    AI Marketing Manager

  • Julia Jaskólska

    Senior Manager, Brand Studio

  • Ronelle Thomas

    Senior Design Producer

  • Jeff Walls

    Brand Design Lead

  • Kelly Galeano Arce

    Senior Brand Designer

  • Janine Anderson

    Sr. Web Marketing Manager

WITH THANKS TO

We're grateful to the following people outside of Zapier, who reviewed early versions of this research and offered feedback:

Femi Agboola, Managing Director, AI & Enterprise Transformation, Citi
Dan Chow, Head of Global IT and Digital Workplace, TikTokUSDS
Chase Darden, HR Business Partner, Follett Higher Education
Cassidy Edwards, Director, AI Enablement and HRBP, Homebase
Sean Glang, Senior Director, Applied AI, Recurly
Lari Hämäläinen, Senior Partner, McKinsey
Alasdair Hay, Senior Solutions Architect, OpenText
Dr. Rebecca Hinds, Head of the Work AI Institute, Glean
Haje Jan Kamps, Senior Operations Director, Editorialist
Ryan Karlin, VP, Product Management, OneTrust
Myra Lane, VP, Business Transformation, Consilio
Udi Ledergor, Chief Evangelist, Gong
Ivan Markman, President and COO, Mediaocean
Chris McDermott, Engineering Manager, Denso
Jeff Mills, Chief Customer Officer, AI, Toptal
Vijay Morampudi, SVP of AI and Innovation, March
Brian Poppe, SVP, Life Insurance Solutions, Mutual of Omaha
Jeremy Renken, Chief Operating Officer, Optspot
Martin Rydder, VP Finance, LiveRamp
Kate Sama, Sr. Director of AI Automation, HubSpot
Timucin Uygungil, Director of Operations, AI, HubSpot
John Yu, Biz Ops and Corp Dev, Paxos

06

Appendix

Population, pricing, sensitivities and boundaries.

How to interpret the findings. This study combines observed activity from real Zapier workflows with measured and modeled costs of running equivalent intended work on the OpenAI and Anthropic products tested. The findings are sensitive to workflow activity, the frequency with which model-⁠led systems check for work, model choice, and account configuration. They describe recurring running costs under the stated assumptions and do not establish equivalent output quality, response times, or total cost of ownership.

How to price one of your own workflows

The method of this study in six steps. Start from the figures in How OpenAI, Anthropic and Zapier were priced; measure the three inputs marked * on your own account.

1 · Checks
How often would the model-⁠led version check for work? Use the builder’s schedule if there is one. If not: 24 a day for replies to people and urgent alerts, 12 for routine heads-⁠ups, once a day for filing and deadlines. Batch work runs on its own schedule.
2 · Empty check*
What does one check cost when it finds nothing? Schedule the task hourly for a few days with no work waiting, then read the billing export. On Claude, this also depends on how many connectors your account has*.
3 · Checks that find work
Using last month’s run times and the proposed checking schedule, count the checking intervals in which at least one run occurred; those are the checks that find work. Note how many items each would handle, since one check may cover several items the original workflow processed separately.
4 · Model-⁠led month*
Monthly running cost = (empty checks × cost per empty check) + (checks that find work × cost per such check). Empty checks equal total checks minus checks that find work. If a check will handle several items, measure its cost at that batch size; the study measured one item.
The worked example uses the study’s ChatGPT estimate of 1.2× empty-⁠check cost for a check that finds work and handles one item. Costs for other products and larger batches may differ.
5 · Deterministic month
(Billable tasks × your task rate) + (what your model provider bills for third-⁠party AI steps, including reasoning tokens)
6 · Compare
Model-⁠led month ÷ deterministic month. Repeat for each product you are considering, and for your busiest and quietest months.

* Measure these on your own account rather than borrowing them: the cost of an empty check, on Claude the connector count, and the cost of a check that finds work at your batch size. Everything else can start from the figures in How OpenAI, Anthropic and Zapier were priced. See the worked example.

Worked example: 20 hourly workflows at two levels of activity

A team’s 20 workflows that watch an inbox, each checking every hour, priced the way the study prices them, on ChatGPT Workspace Agents

Each finds work 40 times a monthEach finds work 600 times a month
Checks a month20 × 720 = 14,40020 × 720 = 14,400
Checks that find nothing13,6002,400
Cost of one empty check0.19 credits × $0.04 = $0.00760.19 credits × $0.04 = $0.0076
Cost of empty checks13,600 × $0.0076 = $103.362,400 × $0.0076 = $18.24
Cost of checks that find work (× 1.2)800 × $0.00912 = $7.3012,000 × $0.00912 = $109.44
Model-⁠led month$110.66$127.68
Deterministic month800 runs × 3 tasks × $0.003 = $7.2012,000 runs × 3 tasks × $0.003 = $108.00
Multiple15×1.2×

Illustrative unit prices from How OpenAI, Anthropic and Zapier were priced, not account bills. One item handled per run, so no per-⁠item amount. The AI step is an AI by Zapier step, which is task-⁠priced, so the deterministic month carries no provider charge. The multiple is the same for one workflow or twenty; only the dollars scale. Claude prices a check as API requests and in this study came out higher. At 20 finds a day a workflow sits in the most active group on the hourly chart, where the median was 2.4×.

Hourly checks: 180 workflows

Median cost multiple for the 180 workflows that check about hourly and have a Zapier charge.

TIMES A DAY THE HOURLY CHECK FINDS WORK

Under 0.5 · 41 workflows
418×
0.5 to 2 · 44 workflows
79×
2 to 4 · 24 workflows
33×
4 to 8 · 9 workflows
7.3×
8 or more · 62 workflows
2.4×

Log scale. Counts are workflows in each group.

Workflows with 20 to 28 checks a day, all connectors, with a Zapier charge (180 workflows; Slack-⁠triggered workflows priced on OpenAI’s channel trigger are left out, since they do not check on a schedule). For each workflow its modeled model-⁠led running cost is divided by its deterministic running cost; each bar is the median of those ratios in its group, which describes the typical workflow without a few expensive ones setting the figure.

Where most checks find nothing, model-⁠led cost 14× to 57× as much

Each cell shows how much the workflows in it cost on model-⁠led relative to deterministic, by how often they run and how often their checks find work.

ALL WORKFLOWS

2,944 workflows · 2.1× overall

FREQUENT

30+ completed runs a month · 1.2× overall

INFREQUENT

Under 30 completed runs a month · 8.3× overall

MOST CHECKS FIND NOTHING

Sentinels: more than half of a workflow’s checks come back empty

FREQUENT14×332 workflows (11% of workflows)24% of the model-⁠led bill
INFREQUENT57×1,020 workflows (35% of workflows)42% of the model-⁠led bill

MOST CHECKS FIND WORK

Fewer than half of checks come back empty

FREQUENT0.62×821 workflows (28% of workflows)25% of the model-⁠led bill
INFREQUENT1.7×771 workflows (26% of workflows)10% of the model-⁠led bill

A check is counted as empty when the modeled schedule wakes the workflow and no work has arrived. Sentinel here means more than half of a workflow’s checks are empty; the infrequent-⁠sentinel cell holds at every threshold tested (50%, 75%, 90% of checks empty), the frequent cells shift as the threshold moves. Model-⁠led average of OpenAI and Anthropic; Zapier at $0.003 a task plus estimated provider charges for third-⁠party AI steps. Figures use the same pricing assumptions as the overall comparison.

2,944 workflows from 360 companies

Of the 375 companies in the first Index panel, 360 had an AI workflow that met this edition’s eligibility rules; their 2,944 workflows are the study.

3,279

AI workflows from the Q2 Index

373 companies; ran in July and August 2026

−373

Webhook-⁠triggered

No app for a scheduled check to look at

−75

Sub-⁠workflows

Started by another workflow

+113

Agent-⁠step workflows

Not among the 3,279; admitted under a separate test (see Methods and disclosures: Agent-⁠step workflows)

2,944

Workflows in this study

Priced on Zapier, ChatGPT and Claude

1,153

Frequent: 30 or more completed runs in the month

1,791

Infrequent: fewer than 30 completed runs in the month

We also checked a wider set: 4,102 AI workflows active in two sample windows in Q3. Using that set would raise the multiples by a few percent to about 10%, and no conclusion would change. About 1,070 newer AI workflows are not included because of how the original sample was drawn.

Cost scope. We priced real run activity, Zapier tasks at $0.003, estimated provider charges for third-⁠party AI steps (including reasoning tokens), and modeled model-⁠led usage. Not priced: seats, build effort, maintenance.

How OpenAI, Anthropic and Zapier were priced

Measured where we could, documented where we could not, modeled in between

Zapier

What we did: counted every run and billable task over 30 days, repriced the tasks at $0.003, and added estimated provider charges for third-⁠party AI steps. Estimated provider charges include 1,000 reasoning tokens a call where the model reasons by default (a chosen figure).

ChatGPT Workspace Agents

What we did: measured what one empty check costs in the product, by connector, and multiplied by each workflow’s checks; a check that finds work costs 1.2× plus a small per-⁠item amount. Measured October 2026, hourly checks: Slack 0.15 credits, email 0.19, a Drive folder compared against a sheet 0.74, all other 0.17; $0.04 a credit; $0.00155 for each further item (one-⁠item wakes). About a quarter of smaller-⁠model workflows priced on the larger model, per the replay; workflows built on the larger model at its rate (9.5× the smaller). Start-⁠up context already cached, as measured in an organization with steady use. Slack-⁠triggered workflows are priced at the lower of the measured channel-⁠mode cost (one billed run per message posted) and the scheduled-⁠check cost.

Claude scheduled tasks

What we did: built a cost model one API request at a time and calibrated it against metered cloud runs. An empty wake is one harness request (the standing context the task loads on every wake: instructions, tools and connectors) plus the task’s own reads and writes (13.6 requests on average). Haiku 5.5 throughout; 1-hour cache writes at 2× input; later runs re-⁠read most of the standing context from cache. Central case: the account’s connectors keep the first request under the 100,000 tokens above which Claude bills the long-⁠context rate (Haiku 5.5; Anthropic pricing page, read Oct 30, 2026). Central case keeps the smaller model throughout: in the measured wake, a larger-⁠model share would reprice every request in the wake, mostly checking. A larger-⁠model share is shown on the sensitivities page.

HOW OFTEN A MODEL-⁠LED WORKFLOW CHECKS (WHERE THE BUILDER STATED NO SCHEDULE)

Reply to a person24 a day
Urgent alert24 a day
Routine heads-⁠up12 a day
Filing1 a day
Deadline1 a day
BatchBuilder’s schedule; once a day where unknown

Builder-⁠stated timing used where it exists (Schedule triggers, digests, waits and delay-⁠until steps, 639 workflows). Who receives the output sets the rest.

Combined figure: OpenAI’s and Anthropic’s monthly dollars averaged with equal weight, divided by Zapier’s for the same workflows.

How the combined figure is calculated

FOR EACH WORKFLOW

Model-⁠led average cost = (ChatGPT cost + Claude cost) ÷ 2

FOR ANY GROUP OF WORKFLOWS

Cost multiple = total model-⁠led cost ÷ total deterministic cost. Deterministic cost = Zapier tasks × $0.003 + provider charges on third-⁠party AI steps.

Because the average is linear, this gives the same group multiple whether the averaging is done per workflow or per group.

Workflow by workflow, 78% cost more on model-⁠led (2,063 of the 2,645 with a Zapier charge; see the workflow-⁠level outcomes chart).

Why we divide totals. The headline multiple divides total model-⁠led spend by total deterministic spend, so it reflects the cost of running the whole portfolio. That gives high-⁠volume workflows more weight, which is why the overall 2.1× sits closer to the frequent 1.2× than the infrequent 8.3×. We also report the share of workflows that cost more on each approach, so you can see the picture workflow by workflow.

In total, model-⁠led cost more in every case we tested

Each scenario changes one assumption from the central 2.1× estimate. Checking frequency and account configuration each change the result substantially.

Larger-⁠model share applied on the Claude column (low / central / high)2.58× / 2.73× / 3.06×
2.58× / 2.73× / 3.06×
Larger-⁠model share on the ChatGPT column (measured floor / ceiling)1.95× / 2.26×
1.95× / 2.26×
Small organization (start-⁠up not cached)2.18×
2.18×
Account with enough connectors that Claude bills its first request at the long-⁠context rate4.43×
4.43×
Claude harness only, at our connector count2.08×
2.08×
Excluding the highest-⁠volume workflows (a defined set, recorded in the research record)3.7×
3.7×
Assigned checking frequency halved / doubled1.3× / 3.7×
1.3× / 3.7×

On the central estimate Zapier is 48 percent of the model-⁠led cost (2.1×); across the rows shown it ranges from 76 percent (1.3×) to 23 percent (4.43×).

What the ranges mean. The 2.0× to 3.2× band covers alternative assumptions about model choice; it runs from the ChatGPT column’s measured floor to both columns at their high larger-⁠model shares together. Checking frequency is an estimated input for most workflows: halving or doubling the assigned checking frequency gives 1.3× to 3.7×. Neither band is a statistical confidence interval, and neither covers every plausible account configuration or combination of assumptions.

All figures price Zapier at $0.003 a task plus the estimated provider charges on third-⁠party AI steps, and use the full study population unless otherwise noted. The checking-⁠frequency row scales each workflow’s modeled checking cost with its assigned checking frequency; builder-⁠stated schedules (639 workflows) are held fixed. The long-⁠context rate applies above 100,000 tokens for Haiku 5.5 (Anthropic pricing page, read October 30, 2026).

What this study does and does not price

PRICED HERE
  • Recurring running cost of a defined workload
  • Zapier: tasks and estimated provider charges
  • ChatGPT: product usage at the published rate card
  • Claude: modeled tokens at list prices
NOT ESTABLISHED
  • Build effort and ongoing ownership
  • Output equivalence beyond the model agreement test
  • Response time: scheduled checks and event triggers run at different speeds.
  • Seats (assumed already held)
  • Separately developed event handlers, custom connector integrations and services that invoke agent APIs (see Boundaries of the finding: Builder scope)
  • The engineered build behind webhooks; Claude Code Routines and Managed Agents (developer products, outside this study’s audience)
  • Quality and reliability
WHAT THE FIGURES ARE

These findings apply to the products, prices and builds we tested in October 2026. They do not show that every model-⁠led workflow costs more, particularly one that starts when work arrives instead of checking on a schedule. Multiples are model-⁠led monthly running cost ÷ deterministic (Zapier) for the same workflows, averaged equally across ChatGPT Workspace Agents and Claude scheduled tasks; Zapier at $0.003 a task plus estimated provider charges. Range across the measured larger-⁠model shares: 2.0× to 3.2×.

Boundaries of the finding

Architecture, equivalence, response times and scope

Architecture or pricing
We applied the same observed workflow structure and monthly activity to implementations on Zapier, OpenAI and Anthropic. The resulting cost difference reflects both how the tested implementations orchestrate work and how each product charges for it. Read it as a comparison of these implementations at October 2026 prices.
Equivalent work
We preserved the workflow steps and intended tasks when constructing the model-⁠led comparisons, and separately tested agreement between the models and the incumbent AI steps. Agreement is an imperfect measure of quality: it does not establish that the implementations would produce equally correct results in production. The study compares modeled running costs for equivalent intended work, without claiming proven equivalence in accuracy or reliability.
Response times
The modeled workflows use checking schedules the products support, with Claude scheduled tasks checking no more often than hourly. Event-⁠triggered workflows can begin when work arrives, so the implementations may differ in response time even when they perform the same steps. No economic value has been assigned to that difference.
Checking and the model
In the cloud products and configurations tested, connector authorization was handled outside the script environment, and inspecting connected services involved model-⁠mediated tool use. That describes the execution paths available in those products during testing. It does not establish that every agent architecture must use a model to detect incoming work.
Build versus buy
The study measures recurring running cost and does not estimate the engineering effort to build and maintain either approach. Recommendations about retries, monitoring, access controls and ownership are decision criteria for leaders to weigh alongside the measured costs, not additional quantified findings.
What model-⁠led means here
We classify orchestration by what controls the work during execution. In this study, all deterministically orchestrated workflows start through software triggers. Most of the model-⁠led workflows instead start a model session on a schedule to check for work. That difference explains much of the cost gap, but a software trigger alone does not make orchestration deterministic. Slack-⁠triggered Workspace Agents start when a message arrives, then use a model to control the work. They remain in the model-⁠led comparison and are priced at the lower of the trigger path and scheduled checks, since a builder would choose the cheaper of two paths on the same product. Workflows with a predefined surrounding sequence and bounded agent steps are classified as deterministic workflows containing agent steps.
Builder scope
We included configurations available through native product features and built-⁠in connectors, including schedules and supported event triggers. We excluded separately developed integrations that require their own deployment or maintenance. AI-⁠generated code running within the tested products was allowed. These boundaries define the configurations compared.

Methods and disclosures

Checking intervals, model choice and ChatGPT measurement

Checking intervals
Where workflows specified timing through Schedule triggers (469), digests (56), waits over five minutes (114) or delay-⁠until steps (13), we used it. For the rest, one analyst assigned checking intervals by how quickly the output would ordinarily be needed: more frequent checks for human-⁠facing work and urgent alerts, less frequent for filing and deadline-⁠related tasks. These assignments are estimates rather than observed behaviour and are the largest source of uncertainty in the modeled cost of infrequent workflows. The four sources can overlap, so their sum is not the count of uniquely scheduled workflows.
Model choice
The share of workflows priced on the larger model was first set from an agreement test on 137 randomly sampled small-⁠model workflows (685 test cases, inputs written by Claude Opus, agreement graded by gpt-4.1, 30 cases spot-⁠checked by the author), then measured on 1,519 workflows: about a quarter need the larger model; the two vendors’ measurements were within two points of each other at the floor, the midpoint and the ceiling. Gemini-⁠incumbent workflows were excluded.
ChatGPT measurement
We did not count the tokens spent setting up each agent, only the tokens spent running it. We did count the tokens an agent uses to work out which values to put where, a step the Zap does without a model.

Claude measurement and items per wake

Claude measurement
Measured in Claude’s cloud scheduled tasks. A bare wake is one request; an empty wake of a real task averaged 13.6 requests. In our tests, scheduled tasks carried every connector on the account, so the size of the first request, and whether it passes the 100,000 tokens above which Claude bills the long-⁠context rate for Haiku 5.5 (Anthropic pricing page, read October 30, 2026), depended on the account’s connector count. In our tests, hourly was the fastest a task could run; the model is set per task.
Items per wake
A wake that finds work was measured handling one item. The ChatGPT column charges 1.2× an empty check for the wake plus the measured $0.00155 for each further item; the Claude column prices items in its request model. How a session’s cost grows past a handful of items per wake is not measured on OpenAI or Anthropic, and some high-⁠volume workflows would need a session to handle 100 or more records in one wake. The cost model assumes that session-⁠start cost can be spread across several items. Our one-⁠item measurements do not validate that assumption at higher volumes, so estimates for workflows handling many items per session are less certain, and more likely understated than overstated.

How we checked the cost estimates

What was measured behind the modeled figures

Working automations measured
OpenAI: 82 Workspace Agents, built from the instructions of 32 workflows (34 instruction sets), plus eight test agents with no customer content. Anthropic: Claude scheduled tasks rebuilt from 32 workflows, plus nine live pilot tasks (eight local, one cloud) on an author’s own accounts.
Sessions measured
ChatGPT: the scheduler fired 191 of 191 scheduled hours across the planted-⁠message rounds; about 3% of runs failed and were not billed. Claude: 63 instrumented cloud sessions, plus about 1,005 wakes in the first week of the live pilots. Both products: empty checks and productive wakes handling one item.
Connectors and workloads covered
Measured empty-⁠check cost on OpenAI: Slack, Gmail, and a Drive folder compared against a sheet; every other connector is priced at a common 0.17-credit estimate. On Claude: accounts with 9 and 11 connectors, which sit either side of the 100,000-token line above which the long-⁠context rate applies to Haiku 5.5 (Anthropic pricing page, read October 30, 2026).
Items per productive session
One, on both products. Productive wakes handling 20 or 100 items were not measured. The OpenAI column prices each further item at a measured $0.00155; the Claude column scales with its request model.
Predicted against measured
Claude: the cost model built before measurement predicted about $0.007 and 3 requests for an empty wake; measurement showed $0.149 and 13.6 requests, and the model was refitted to the measured values. ChatGPT: no prior model; the measured per-⁠connector costs are used directly, with a productive wake at 1.2× an empty one (measured 15 to 20% more).
Calibration, not validation
The Claude model was fitted to the sessions it was measured on; no separate set of sessions was held back to test it. The ChatGPT figures are measurements applied as unit costs. Modeled costs have not been compared with billed costs on a customer account.

Methods and disclosures

Pricing and model choice

Vendor cost validation
For ChatGPT we measured the cost of scheduled checks in the product across selected connectors. For Claude we built a request-⁠level cost model and calibrated it against metered cloud runs. We applied those measurements to the observed activity of every eligible workflow, so the portfolio results are modeled estimates, not bills from running each customer workflow on each vendor.
Unmeasured connectors
Empty-⁠check cost was measured for selected connectors; a common 0.17-credit estimate was applied to connectors without a dedicated measurement, so that no check was priced as free. The estimate adds uncertainty for integrations whose retrieval or context needs differ materially from the measured cases.
Smaller model throughout
The central Claude estimate keeps the smaller model throughout. In the measured wake, moving a workflow to the larger model reprices every request in that wake, most of which is checking rather than the work, and the replay’s quality gap on the smaller model was small. The sensitivity rows show the result with a larger-⁠model share applied. The central estimate is a lower-⁠cost scenario, not a validated quality-⁠matched configuration for every workflow.
The 60 / 32 / 8 split
The shares of model-⁠led spending going to checks that find nothing, session start-⁠up when work is found, and the work itself are modeled allocations. They use observed workflow activity and the modeled per-⁠check costs, including the assumption that a check that finds work costs 1.2× an empty one before further items. They are not line items reported by the vendors.
Model-⁠led construction
The model-⁠led implementations used scheduled tasks with explicit instructions, rules and a record of previously handled work, configured to avoid unnecessary model activity within the products’ capabilities. The comparison does not cover every custom-⁠coded implementation, including separately engineered event handlers or connector integrations.
One assumption at a time
Sensitivities move one assumption at a time from the central 2.1× while holding the others constant, which makes each effect easy to read but does not capture interactions among checking schedules, model choice, caching, connector count and items handled per session. The combined effect may be larger than the individual rows show. The share levels behind the model-⁠choice rows are recorded in the research record. The checking-⁠interval row, on which the main results are most conditional, halves and doubles the assigned checking frequency for the workflows without a builder-⁠stated schedule.
Zapier task price
We price Zapier at USD 0.003 a task, the monthly-⁠billed rate of the highest published Team tier (2M tasks a month). Smaller plans pay more per task and the multiples fall. Because the deterministic cost also carries estimated provider charges, the multiple is close to, but not exactly, inversely proportional to the task rate.

Population and interpretation

Representativeness
The study uses durable AI workflows from leading Zapier adopters because those accounts show ongoing activity rather than isolated experiments. The results describe that population and should not be taken to represent all businesses using AI.
360 of 375 companies
We started from the 375 companies in the first Index panel and kept every AI workflow that ran in both July and August 2026 and could be run by a scheduled agent, plus 113 agent-⁠step workflows admitted under a separate activity test (next page): 2,944 workflows from 360 companies. Two companies fell out before the eligibility test: one was no longer in the cohort table when the population was drawn on September 4, 2026 (the table drops paused workflows); the other was in the cohort table, but none of its AI workflows ran in both July and August 2026. The 373 webhook-⁠triggered workflows excluded under Population and scope are a separate count; the match with the 373 companies is coincidence.
The 30-run boundary
Thirty completed runs a month, about one a day, is a practical reporting boundary between frequent and infrequent workflows, not a statistically estimated breakpoint. The economics also vary substantially within each group, particularly with the number of checks that find no work.
Role comparisons
Each workflow was assigned one primary role from its configuration and the instructions supplied to its AI steps. The 609 workflows without a clear primary role remain in the overall findings but are omitted from the role comparison. Clerk results are sensitive to a concentration of unusually high-⁠volume workflows, so comparisons across roles should be read with that in mind.
Large multiples
Cost multiples become very large when the deterministic version of a workflow costs very little. They show how the economics change as scheduled checks become less productive; they do not imply that every workflow with a large multiple represents substantial dollar savings. Actual spending should decide which workflows merit redesign.
The hourly chart
The hourly chart in the appendix covers the 180 workflows that check 20 to 28 times a day and have a Zapier charge, grouped by how often a check finds work. Each bar is the group median, so a few very expensive workflows do not set the figure; the 41 least-⁠active workflows, at 418×, are shown with the rest. Slack-⁠triggered workflows priced on OpenAI’s channel trigger are left out because they do not check on a schedule; they remain in the study’s aggregate calculations.
Agent-⁠step workflows
The 113 agent-⁠step workflows are not among the 3,279. They were admitted under a separate test (ran in both of two 31-day windows between July 15 and September 14) because the cohort’s definition of an AI step did not count the Run Agent step. A further 22 workflows inside the 2,831 that already had agent steps were repriced in place, not added. Including the 113 raises the overall multiple by less than 0.1×. All were priced through the same engine as every other workflow, with each tool call priced as one further model request per item handled; their checking cadence follows the same rule as the rest. Their provider charges are not recorded, so their deterministic cost is tasks only.

Notices and disclaimers

AI Workflow Index, Q3 2026 · Zapier Research Lab

Copyright
© 2026 Zapier, Inc. All rights reserved.
Trademarks and no endorsement
Zapier, the Zapier logo and other Zapier marks are trademarks of Zapier, Inc. ChatGPT and OpenAI are trademarks of OpenAI; Claude and Anthropic are trademarks of Anthropic, PBC. Other product and company names are the trademarks of their respective owners. Their use does not imply affiliation with, sponsorship by, or endorsement by those companies.
Informational only, not advice
This report is provided for general informational purposes only. It is not legal, security, compliance, financial or purchasing advice and should not be relied on as such. Architecture, vendor and deployment decisions remain the responsibility of the reader and their organization.
Data basis
All statistics are aggregated and de-⁠identified. They are derived from configuration and usage metadata: workflow structure, step types, connected apps, and run and task counts. The instructions builders configured in AI steps were read to assign AI roles, to rebuild those steps as test automations on the products compared, and to test them on OpenAI and Anthropic models with synthetic inputs written from each workflow’s configuration. They are not reproduced here. No customer content processed by a workflow at runtime was used, and the aggregated figures do not identify any individual company or person.
Modeled-⁠figure qualifier
Model-⁠led costs are modeled estimates. They were measured on Zapier’s own accounts and applied to the structure and run history of live customer workflows. Tests used synthetic inputs, except nine pilot tasks rebuilt from customer workflows (eight local and one cloud), which ran read-⁠only on an author’s own work accounts to measure checking costs on live traffic; their outputs were reduced to counts and redacted paraphrases. No customer workflow was rebuilt with data it processed at runtime. Results will vary with workflows, volumes, accounts and configuration.
Pricing inputs
Pricing inputs are public list prices for Zapier, OpenAI and Anthropic products as read in October 2026. They are illustrative, are not guaranteed rates for any plan, and may have changed since. Product versions and capabilities are as tested in October 2026.