What an AI Agent Costs to Run Per Month

AI agent running cost per month is not the model bill. The model bill is often the smaller number.

The full cost is tasks times calls times tokens times price, plus the human time spent reviewing and monitoring.

Below is the formula, a worked example on made-up numbers, and the questions to ask any vendor. I use placeholder prices on purpose. Model prices change often, and you should plug in the current ones.

Build cost versus run cost

Two different budgets get mixed up.

Build cost is the one-time work to design, build, test, and deploy the agent. I cover that in what AI agent development costs.

Run cost is what you pay every month after launch. It does not stop. It scales with volume, and it includes people.

CFOs tend to approve the build number and then get surprised by the run number. Put both in the same business case. For the full ROI math, see how to calculate AI agent ROI, and for the finance view of the whole investment, the CFO guide to AI agent investment.

The AI agent running cost per month formula

Run cost per month has five parts:

Model cost    = tasks x calls per task x (input tokens x input price
                                         + output tokens x output price)
Retry cost    = model cost x retry rate
Review cost   = tasks x share reviewed x minutes per review x hourly rate
Monitoring    = hours per month x hourly rate
Tooling       = hosting, logging, search, and vendor fees

Run cost      = model + retry + review + monitoring + tooling

Then the number finance cares about:

Cost per completed task = run cost / tasks completed correctly

Use “completed correctly,” not “attempted.” An agent that runs 10,000 times and gets 8,500 right has a cost per completed task based on 8,500.

Why agents make many calls

A chatbot answers a question in one call. An agent runs a loop. It reads the request, picks a tool, calls it, reads the result, decides what to do next, and repeats.

Every step is a model call. And each call usually carries the whole conversation so far: the instructions, the tool definitions, and the results of earlier steps. Anthropic describes an agent in a loop as generating more and more data that might matter for the next turn.

That is why cost grows faster than you expect. Anthropic’s own data, shared in its multi-agent research write-up, is that agents typically use about 4 times more tokens than chat interactions, and multi-agent systems about 15 times more. Treat that as one vendor’s measurement on its research product, not a law. But the direction holds. In its guide to building agents, Anthropic says agentic systems often trade latency and cost for better task performance, and that autonomy means higher costs and the potential for compounding errors.

Here is the effect on one made-up task, using placeholder prices of $2 per million input tokens and $10 per million output tokens. These are not any real model’s prices.

	Chatbot answer	Agent task
Model calls	1	6
Input tokens per call	1,500	8,000
Output tokens per call	300	600
Total input tokens	1,500	48,000
Total output tokens	300	3,600
Cost at placeholder prices	$0.006	$0.132

That is about 22 times more per task. Your ratio will differ. What matters is that the multiplier comes from three places at once: more calls, bigger context per call, and more output.

Review and monitoring time

Most agents in business settings do not run unattended. A person approves, corrects, or handles exceptions. That time is real cost, and in my experience it often beats the model bill.

Count three kinds of human time:

• Task review: the share of tasks a person checks, times minutes per check.
• Exception handling: the cases the agent hands off, which a person finishes.
• Monitoring and maintenance: someone watches quality, reads failures, updates prompts and tools, and re-tests when the model or a source system changes.

The first two shrink as the agent improves. The third never hits zero. See AI agent maintenance after deploy if you want the detail on what upkeep involves.

Worked example with made-up numbers

Say a 40-person finance operations team uses an agent to triage incoming invoices. Every number below is hypothetical, so swap in yours.

Assumptions

• 10,000 tasks per month.
• 6 model calls per task, 8,000 input and 600 output tokens each.
• Placeholder prices: $2 per million input tokens, $10 per million output tokens.
• 70% of input tokens are served from a cache at 10% of the normal input price. (A placeholder discount. Vendors publish theirs.)
• 15% extra cost for retries and failed runs.
• 20% of tasks get a 3-minute human review at $40 per hour loaded cost.
• About 4 hours per week of monitoring at $90 per hour.
• $400 per month for hosting, logging, and tools.
• 85% of tasks end up correct.

The math

Line	Calculation	Monthly
Input tokens	10,000 x 6 x 8,000 = 480M
Output tokens	10,000 x 6 x 600 = 36M
Input cost, uncached	480M x $2 per M	$960
Input cost, with cache	480M x (30% x $2 + 70% x $0.20) per M	$355
Output cost	36M x $10 per M	$360
Model cost (with cache)		$715
Retries at 15%		$107
Human review	2,000 tasks x 3 min = 100 hours x $40	$4,000
Monitoring	about 17 hours x $90	$1,560
Tooling		$400
Total run cost		about $6,780

Per task: about $0.68. Per correct task: about $0.80, since 8,500 of 10,000 were right.

Look at the shape. The model is about 11% of the total. Human review is about 59%. If you only budget for tokens, you miss most of the cost.

Is $0.80 good? It depends on what the task costs today. If a person spends 12 minutes on each invoice at the same $40 hourly rate, that is $8 per task. Then the agent costs a tenth as much, even with review. That is a hypothetical comparison, but it shows the kind of math that makes the case.

Levers that cut cost

Ranked by what I’d try first:

1. Cut the review share as accuracy proves out. Moving review from 20% to 10% saves $2,000 a month in this example, more than any token trick. Do it only with evidence from your test set.
2. Cache repeated context. Instructions and tool definitions repeat on every call. Anthropic says its prompt caching can cut costs by up to 90% and latency by up to 85% for long prompts. That is a vendor claim for best cases, so measure your own.
3. Batch work that can wait. Anthropic’s Message Batches API offers a 50% discount for asynchronous work returned within 24 hours. Overnight jobs fit.
4. Trim context. Drop old tool results, summarize earlier steps, and pass only what the next step needs.
5. Use a smaller model for easy steps. Routing, classification, and extraction often do not need the largest model.
6. Cap the loop. Set a maximum number of steps per task. Anthropic notes that stopping conditions such as a maximum number of iterations are common practice to maintain control.

Spend caps and budgets

An agent with no limits can spend your money for you. A bug, a loop, or a malicious input can all do it. OWASP calls this unbounded consumption, including “denial of wallet,” and lists rate limits, quotas, timeouts, and monitoring as mitigations.

Put caps at four levels:

• Per task: maximum steps, maximum tokens, and a timeout. When a task hits a cap, it stops and routes to a person.
• Per day: a ceiling on total model spend, with an alert at 50% and 80%.
• Per project: separate API keys or workspaces per agent so one runaway does not drain the rest. Providers support this. Anthropic’s rate limits docs describe organization spend limits and custom spend and rate limits per workspace.
• Per user or source: rate limits so one caller cannot trigger thousands of runs.

Also keep a kill switch someone can reach in under a minute.

Recurring costs people forget

Check this list against your budget:

• Re-testing after a model version change or retirement.
• Evaluation runs. Each regression test is itself a batch of model calls.
• Log storage, especially if you keep full traces for audit.
• Vector database or search hosting, if the agent retrieves documents.
• Prompt and tool maintenance as source systems change.
• Security reviews and access audits.
• Vendor platform fees and per-seat charges on top of usage.
• Incident time when something breaks.
• Human review that never goes down because nobody trusts the agent enough to cut it.

What to ask vendors

Whoever builds or sells you an agent, ask these:

1. What is the expected cost per completed task, and what assumptions sit behind it?
2. How many model calls does a typical task make, and what is the worst case?
3. What caps exist per task, per day, and per project? Who can change them?
4. How is pricing built: usage, per seat, per outcome, or a flat fee? Some products now charge per resolved outcome, as in Intercom’s Fin. Ask exactly what counts as an outcome and what happens when a customer reopens the case.
5. What happens to my bill if the vendor changes the underlying model?
6. What monitoring ships by default, and what does it cost?
7. Can I see a per-task cost report from the first week?

A vendor who cannot give you a per-task number has not measured it.

Bottom Line

Run cost is tasks x calls x tokens x price, plus review, monitoring, retries, and tooling.

Agents cost more than chatbots because they loop, and each loop resends context.

Humans in the loop are often the biggest line item.

Next step: build this table in a spreadsheet with your own volumes, then measure real token use in the first pilot week and replace every placeholder.

FAQ

What does it cost to run an AI agent? It depends on volume, calls per task, model choice, and how much human review you keep. Add model fees, retries, review time, monitoring, and tooling. In my made-up example, a 10,000-task month cost about $6,780, and only about 11% of that was the model.

Why do agents cost more than chatbots? A chatbot makes one call per answer. An agent runs a loop of calls, and each call carries growing context. Anthropic reports agents typically use about 4 times more tokens than chat, and multi-agent setups about 15 times more. Your multiplier depends on the design.

What is cost per completed task? Total monthly run cost divided by the number of tasks finished correctly. It is the number to compare against your current manual cost. Use correct completions, not attempts, so failed runs raise the number instead of hiding.

How do I cap AI agent spend? Set limits at four levels: per task (steps, tokens, timeout), per day (with alerts), per project (separate keys or workspaces), and per caller (rate limits). Add a kill switch. Review caps monthly against actual use.

What recurring costs are easy to forget? Evaluation re-runs after model changes, log storage, search or vector hosting, prompt and tool maintenance, security reviews, vendor platform fees, and incident time. The largest hidden cost is usually human review that never shrinks.

Should I budget with current model prices? Use current prices from the provider’s pricing page at the time you plan, and recheck them each quarter. Prices change, and so do cheaper model options. Keep prices as input cells in your spreadsheet so one edit updates the whole model.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect