n8n AI Agents vs Zapier Agents vs Make AI Agents: Real Agent Automation Cost in 2026
The short answer: for an AI agent, the automation platform is usually the smaller bill and the model is the larger one, and a run that fails still costs money. At 10,000 agent runs a month, a simple 2-tool lookup agent costs about $24 to $27 per 1,000 successful runs on n8n Cloud or Make, a 5-tool business agent about $88 to $94, and a 10-tool autonomous agent with retries about $453 to $527. Zapier Agents cannot be priced at those volumes from public rates: its Pro plan includes 1,500 activities a month, which under the illustrative activity counts below covers only about 500, 250 and 94–100 runs a month for the three shapes, so every requested scenario (1,000, 10,000 and 100,000 runs) needs an Enterprise contract, and that price is QUOTE-ONLY / UNKNOWN. The model cost is calculable on its own: about $19, $82 and $425 per 1,000 successful runs. Failure changes the denominator: at an 85% success rate, every successful run carries 18% overhead, and at 65% it carries 54%. Article 1 covers traditional workflow billing. This article covers agents.
How each platform counts an agent run
n8n (REPORTED). The AI Agent node is free at the platform level, and the whole run is one execution regardless of how many LLM and tool calls happen inside it; model tokens are billed by the model provider. Polling triggers and failed runs still count as executions. Cloud Starter (€20 a month annual, 2,500 executions), Pro (€50, 10,000), Business (€667, 40,000); self-hosted Community Edition is free software with unlimited executions.
Zapier Agents (OFFICIAL, Zapier Agents plans). Free: 400 activities a month. Pro: 1,500 activities a month at $400 a year ($33.33 a month). Enterprise: a custom activity allowance at custom pricing. A single run is capped at 10 activities on Free and 40 on Pro and Enterprise. An activity is consumed by trigger use, a knowledge-source lookup, an action, a web browse or scrape, a web search, and certain direct agent interactions. Agent activity usage is separate from normal Zapier task usage, so a Zaps task plan does not fund agent runs. No Enterprise per-activity rate was public (QUOTE-ONLY / UNKNOWN), and none is assumed here. The Zaps and MCP task route is priced separately in the sidebar below and is not Zapier Agents pricing.
Make (REPORTED). AI Agents run as modules inside scenarios, using either Make's built-in model or a customer's own OpenAI or Anthropic key. Make replaced operations with credits in August 2025: one credit for a standard module, more for AI steps, adjusted again in November 2025. Core is about $9–$12 for 10,000 credits, Pro about $16–$21 and Teams $29–$38. The credit cost of the agent module itself was not published in the sources reviewed.
Agent shapes and assumptions
All are ILLUSTRATIVE. The reference model is GPT-6 Sol at Standard short-context pricing, $2.00 per million input tokens and $10.00 per million output tokens; a cheaper model at $0.50 and $1.50 cuts the model line by about 78%.
| Shape | LLM calls | Tokens per run (in / out) | Tool calls | Retry factor | Success rate | Human approval |
|---|---|---|---|---|---|---|
| A: 2-tool lookup | 3 | 6,000 / 600 | 2 | none | 95% | none |
| B: 5-tool business workflow | 7 | 25,000 / 2,000 | 5 | none | 85% | 20% of runs, 3 minutes |
| C: 10+ tool autonomous with retries | 15 | 90,000 / 5,000 | 12 | 1.2 on calls | 65% | 40% of runs, 5 minutes |
Model cost per run: $0.018 (A), $0.070 (B), $0.276 (C, including the 1.2 retry factor), which is $19, $82 and $425 per 1,000 successful runs after the success rates. Platform units per run: Make 8, 11 and 20.4 credits (one for the trigger, one per tool call, five for the agent step, ILLUSTRATIVE) at the Core rate of $0.0009 a credit; n8n 1, 1 and 1.2 executions.
Formulas: cost per 1,000 successful runs = (platform cost + runs × model cost per run) ÷ (runs × success rate) × 1,000. A failed run is not a business outcome: it consumes tokens and platform units, and only successes are counted in the denominator.
Zapier Agents: activities per run and the Pro allowance
Activity counts per run are ILLUSTRATIVE and are based on billable actions: the trigger plus each tool or action call, with retry overhead on shape C.
| Shape | Activities per run (ILLUSTRATIVE) | Runs the Pro allowance covers (1,500 ÷ activities) | Activities at 1,000 runs | At 10,000 runs | At 100,000 runs |
|---|---|---|---|---|---|
| A: trigger + 2 tools | 3 | about 500 | 3,000 | 30,000 | 300,000 |
| B: trigger + 5 tools | 6 | about 250 | 6,000 | 60,000 | 600,000 |
| C: trigger + 12 tools with retries | 15–16 | about 94–100 | 15,000–16,000 | 150,000–160,000 | 1.5–1.6 million |
Every scenario in this article exceeds the 1,500 activities on Pro by at least a factor of two at 1,000 runs, so the Zapier Agents platform price for 1,000, 10,000 and 100,000 runs is QUOTE-ONLY / UNKNOWN. All three shapes stay under the 40-activity per-run cap on Pro and Enterprise, and shape C exceeds the 10-activity cap on Free. The model cost above still applies in any case.
Monthly cost at 10,000 runs
| Shape | Zapier Agents platform | Make (Core rate) | n8n Cloud | n8n self-hosted (infrastructure + labor) | Model cost |
|---|---|---|---|---|---|
| A | QUOTE-ONLY / UNKNOWN (30,000 activities) | $72 | $50 | $30–$205 | $180 |
| B | QUOTE-ONLY / UNKNOWN (60,000 activities) | $99 | $50 | $30–$205 | $700 |
| C | QUOTE-ONLY / UNKNOWN (150,000–160,000 activities) | $184 | $667 (12,000 executions exceed Pro's 10,000) | $30–$205 | $2,760 |
Total cost per 1,000 successful runs
| Shape and volume | Zapier Agents | Make | n8n Cloud | n8n self-hosted |
|---|---|---|---|---|
| A, 1,000 runs | Platform QUOTE-ONLY / UNKNOWN; model $19 | $28 | $40 | $32–$216 |
| A, 10,000 runs | Platform QUOTE-ONLY / UNKNOWN; model $19 | $27 | $24 | $22–$41 |
| A, 100,000 runs | Platform QUOTE-ONLY / UNKNOWN; model $19 | $27 | $68 | $20–$22 |
| B, 1,000 runs | Platform QUOTE-ONLY / UNKNOWN; model $82 | $94 | $106 | $96–$302 |
| B, 10,000 runs | Platform QUOTE-ONLY / UNKNOWN; model $82 | $94 | $88 | $86–$106 |
| B, 100,000 runs | Platform QUOTE-ONLY / UNKNOWN; model $82 | $94 | $137 | $83–$86 |
| C, 1,000 runs | Platform QUOTE-ONLY / UNKNOWN; model $425 | $453 | $455 | $443–$712 |
| C, 10,000 runs | Platform QUOTE-ONLY / UNKNOWN; model $425 | $453 | $527 | $429–$456 |
| C, 100,000 runs | Platform QUOTE-ONLY / UNKNOWN; model $425 | $453 | $496 | $426–$429 |
The n8n Cloud 100,000-run figures include the reported €4,000 overage block per 300,000 executions above the Business plan's 40,000, treated at parity for simplicity; a negotiated Enterprise contract would likely price differently (QUOTE-ONLY / UNKNOWN). Make figures use a linear Core credit rate and are a floor: Make's volume ladder was not captured. Self-hosted figures reuse the Article 1 infrastructure and maintenance assumptions ($12, $30 and $95 of infrastructure plus $175 of labor a month). Make and n8n totals include the model cost.
Platform versus model contribution
At 10,000 runs a month for shape B on Make, the platform is $99 and the model $700, so the model is 88% of the total. On n8n Cloud, the platform is $50 and the model is 93%. For shape C on n8n Cloud, the model is $2,760 against $667 of platform (81%). Choosing a cheaper model changes these totals more than choosing between Make and n8n in most cases. On Zapier Agents, the platform share cannot be computed without an Enterprise quote.
Retry and failure overhead
| Shape | Success rate | Cost multiplier per successful run | Retry factor on model and platform units |
|---|---|---|---|
| A | 95% | 1.05 | none |
| B | 85% | 1.18 | none |
| C | 65% | 1.54 | 1.2 |
In shape C, a 65% success rate and a 1.2 retry factor together make each successful run cost about 1.85 times the cost of a single clean attempt on the model line (1.2 ÷ 0.65). On any platform, the model tokens spent before a failure are still billed.
Break-even: human minutes saved
Formula: minutes needed = cost per successful run ÷ ($40 ÷ 60), with $40 an hour (ILLUSTRATIVE). At 10,000 runs a month:
| Shape | Make | n8n Cloud | Zapier Agents |
|---|---|---|---|
| A | 0.04 minutes | 0.04 | QUOTE-ONLY / UNKNOWN (model alone: 0.03) |
| B | 0.14 | 0.13 | QUOTE-ONLY / UNKNOWN (model alone: 0.12) |
| C | 0.68 | 0.79 | QUOTE-ONLY / UNKNOWN (model alone: 0.64) |
The break-even is a fraction of a minute for every shape. The human approval step costs more than the machine: at $40 an hour, shape B's approvals average 0.6 minutes per run ($0.40) and shape C's 2.0 minutes ($1.33), which is about 4.3 times the platform and model cost of a successful shape B run on Make ($0.094) and about 2.9 times that of shape C ($0.453). Reviewing agent output, not running it, becomes the dominant cost as soon as approvals are required.
Sidebar: the Zapier Zaps and MCP task route (not Zapier Agents pricing)
A different architecture calls tools through Zapier's MCP server from an external agent, billed in ordinary Zap tasks. Zapier's documentation states that each successful MCP tool call consumes two tasks and failed calls consume none (OFFICIAL). Task tiers are those in Article 1 (Professional: 750 tasks $19.99; 5,000 $73.50; 10,000 $103; 50,000 $349; 100,000 $599; 500,000 $1,799; 2,000,000 $5,999, REPORTED). This route is an alternative automation design, and it does not price the Zapier Agents product. Tasks per run are 4, 10 and 28.8 (two per tool call, times the 1.2 retry factor for shape C).
| Shape | Monthly platform at 10,000 runs | Per 1,000 successful runs at 1,000 / 10,000 / 100,000 runs, including model cost | Break-even minutes at 10,000 runs |
|---|---|---|---|
| A | $349 (40,000 tasks) | $96 / $56 / $38 | 0.08 |
| B | $599 (100,000 tasks) | $204 / $153 / $153 | 0.23 |
| C | $1,799 (288,000 tasks) | $962 / $701 / QUOTE-ONLY / UNKNOWN (2.88 million tasks exceeds the listed tiers) | 1.05 |
On this route the model is 54% of the total for shape B at 10,000 runs, and failed MCP tool calls consume no tasks, which lowers the platform side of the failure overhead. Model tokens spent before a failure are still billed.
Sensitivity
- Model price. Swapping to the cheaper reference model lowers the model line by about 78%, taking shape B on Make from $94 to about $30 per 1,000 successful runs.
- Success rate. Each 10 points of success rate moves cost per successful run by 12–15% around the modeled values.
- Agent-step credits on Make. The agent module was assumed at five credits; at 20 credits, shape B rises from 11 to 26 credits per run.
- Plan cliffs on n8n. Shape C crosses Pro's 10,000-execution cap at 10,000 runs and jumps to Business.
- Zapier Agents activities per run. At 3, 6 and 15–16 activities per run, Pro's 1,500 activities cover about 500, 250 and 94–100 runs; a shape that used 40 activities per run (the per-run cap) would cover 37 runs.
Budgeting traps
- Counting runs instead of outcomes. A failed run costs tokens and platform units and delivers nothing.
- Two meters on Zapier. Agent activities and Zap tasks are billed separately, so a Zaps plan does not cover agent runs.
- Activities are not runs. A single Zapier Agents run consumes an activity for every trigger, lookup, action, search and browse.
- Tool-call multipliers on the MCP route. Two tasks per MCP tool call means a 12-tool agent burns about 29 tasks per run with retries.
- Bring-your-own-key. Model charges move to a separate provider invoice.
- Approvals and review labor. These are outside every platform bill.
What to ask before you buy
Instrument 100 real agent runs, count LLM tokens, tool calls, retries and failures, and price them on each platform with the current unit definitions. Ask Zapier for an Enterprise activity price at your volume, and ask how many activities, credits or executions one agent run consumes and what happens to failed runs.