Why Do AI Agents Need Tokens, and What Does That Mean for Your Costs?

Straight answer: AI agents need tokens because every single thing they do, reading a document, thinking through a step, calling a tool, checking their own work, gets converted into text and billed by the chunk. An agent doing a five-step task can use ten to fifty times more tokens than a chatbot answering the same question in one go, which is why your bill from an "agentic" workflow can be shockingly bigger than you expected even though the task itself looks simple on paper.

What a token is, in plain terms

A token is a chunk of text, roughly three quarters of a word in English. "Invoice" is one token. "Reconciliation" might be two or three. Every word you type in, every word the model reads, every word it thinks through, and every word it writes back out gets counted and charged.

With a normal chatbot, you send a prompt, it sends back an answer, and that's the whole transaction. One input charge, one output charge, done. That's the model most people still have in their head when they think about AI costs, and it's the wrong model once you start using agents.

Why agents are a different animal from chatbots

An AI agent isn't answering one question. It's completing a task, which usually means several rounds of thinking, checking, and acting before it gives you anything back. If you want the full picture of what separates an agent from a chatbot, I wrote about that in what an AI agent is, but here's the short version for cost purposes: every one of those internal rounds costs tokens, and none of them are visible to you as the user.

Say you ask an agent to "check our top five competitors' pricing pages and summarise any changes since last month." Behind the scenes it might:

• Plan the steps it needs to take (tokens)
• Search or open each competitor's website (tokens for reading each page)
• Decide whether the page loaded correctly or needs retrying (tokens)
• Compare what it found to what it remembers from before (tokens, and this is where memory comes in, which I covered in whether AI agents remember between tasks)
• Write a draft summary, then check its own draft for accuracy (tokens, twice)
• Format the final answer for you (tokens)

That's six or seven token-consuming steps for what looks, from where you sit, like one request. And each of those steps can fail and retry, which multiplies the cost again.

The afternoon I watched an agent burn eleven pounds on a job that should have cost pennies

I tested this a few months back, not in theory, on my own laptop with my own card attached. I built a simple research agent using Claude to pull pricing and positioning from fifteen competitor sites for a client audit. The task should have taken maybe two thousand tokens in and two thousand out, call it fifteen or twenty pence at typical rates.

One site had a cookie banner the agent couldn't get past cleanly. Instead of giving up, it tried again. And again. It re-read the same blocked page forty-one times before it decided to move on, each attempt logging the full page content plus its own reasoning about why the previous attempt had failed. By the time the job finished, that one stuck page alone had used more tokens than the other fourteen sites combined. The final bill for that single afternoon of testing came to just over eleven pounds, for a task that on a good run costs about twenty pence.

Nobody warns you about this bit. The demos you see of agents working are always the clean runs. The real cost sits in the messy runs, the retries, the loops, the pages that don't load, and those happen far more often than any vendor pitch lets on.

The maths, step by step, with real numbers

Let's put actual figures against this so it's not abstract. Using pricing that's typical for a capable model like Claude 3.5 Sonnet or GPT-4o class models, you're looking at roughly $3 per million input tokens and $15 per million output tokens for the higher-end models, or about $0.15 to $2.50 per million for the cheaper, faster ones.

Now walk through a common small business task: an agent that monitors your inbox, drafts replies, and flags anything urgent.

• Reading one email and its thread history: roughly 500 to 2,000 tokens depending on length
• The agent's internal reasoning about tone, urgency, and what to say: another 500 to 1,500 tokens, invisible to you
• Drafting the reply: 200 to 600 tokens
• Checking the draft against your house style or a template: another 500 to 1,000 tokens

Add that up and one email can use 2,000 to 5,000 tokens once you count all the thinking steps, not just the visible reply. Multiply that by fifty emails a day, five days a week, and you're at roughly 500,000 to 1.25 million tokens a week just for a modest inbox agent. At blended pricing that's somewhere between $5 and $20 a week for one simple agent doing one simple job. It sounds small until you're running six or seven agents across sales, support, and content, and the invoices start stacking on top of each other.

The uncomfortable bit vendors don't spell out

Here's the part that doesn't get said out loud often enough. A lot of the "AI agent" software you can buy off the shelf isn't charging you for tokens at all, it's charging you a flat monthly seat fee, and pocketing the gap between what you pay and what the underlying model costs them. That gap is frequently enormous. A tool charging you £79 a month per user might be spending 60p to £2 in actual token costs on a typical light user, and the rest is margin for the interface, the support, and the sales team.

That's not a scandal, software has always worked that way, but it means the "per seat" pricing you're shown tells you almost nothing about what the AI itself is doing or costing underneath. It's designed to be easy to budget, not to be honest about usage. If you want to understand what you're paying for, ask the vendor directly which model they use and roughly how many tokens a typical task consumes. Most sales reps have never been asked this and won't have an answer ready, which tells you something on its own.

The flip side is also true and just as uncomfortable: if you build your own agents directly on an API instead of buying a wrapped product, you'll pay far less per task, but you carry all the risk of runaway loops like the one I described above, with nobody else absorbing the cost of a bad retry.

Reasoning models make this worse, not better

The newer "reasoning" models, the ones that pause and think before answering, are brilliant for accuracy and worse for your bill. They generate long chains of internal reasoning tokens that you never see and typically aren't shown in the chat window, but you're billed for every one of them. On a complex task, the invisible reasoning can be three or four times longer than the visible answer. This is one of the recurring themes I've flagged in my weekly AI news roundups, because pricing models keep shifting as these reasoning features roll out more widely, and last week's assumptions about cost don't always hold this week.

How to keep agent costs under control

None of this means avoid agents. It means budget for them like you'd budget for any other business tool with variable running costs, not a flat subscription. Here's what I do now, after that eleven pound afternoon taught me the lesson:

• Set a maximum step count. Cap how many times an agent can retry or loop on a single task, five or six attempts, then it stops and flags the failure to a human instead of grinding on.
• Use a cheap model for routing, a good model for output. Let a small, fast model decide what needs doing, and only hand the expensive reasoning model the parts that need judgment.
• Log token use per task, not just per month. A monthly total tells you nothing useful. Knowing that the inbox agent costs 4p per email and the research agent costs £2 per report tells you where to trim.
• Set a hard budget alert. Most API providers let you cap monthly spend or get alerted at a threshold. Use it, because the first time you find out about a runaway loop should not be when the invoice lands.
• Test on messy real data before you trust it on live tasks. Clean demo data never shows you the cookie banners and broken pages that cause the expensive retries.

If your business is scaling past a couple of simple agents and the token maths is starting to feel like guesswork, that's usually the point where it's worth getting a proper cost audit done rather than continuing to trial and error your way through it. I go through exactly this kind of budgeting with clients, and if you want a sense of what that support typically costs, I've broken it down honestly in what an AI consultant charges.

Where this bites hardest in practice

The businesses that get caught out are almost always the ones running agents on customer-facing work, support inboxes, chat widgets, anything with unpredictable input, because you don't control how long or messy the conversation gets. If you're weighing up an AI-powered support tool, this is exactly why I'd tell you to ask about token usage before you sign anything, a point I go into more in what to look for in an AI customer service platform. A customer who won't stop rephrasing the same question five different ways can quietly cost you fifty times more than a customer who asks once and gets a clean answer, and no seat-based pricing plan will warn you when that happens.

If you're building your own agents rather than buying a packaged tool, worth knowing that the frameworks available now, including the approach I walked through in building agents with Claude, all give you some level of control over step limits and tool call budgets. Use that control. It's the difference between a tool that costs what you planned and one that quietly costs three times more by the end of the month.

If you build with language models, you can contribute an LLM article on my blog.

If you write about decentralised tech, see my write for us web3 page.

Frequently asked questions

Do AI agents cost more than regular chatbot subscriptions?

Often yes, because agents make multiple internal calls to complete one task while a chatbot subscription usually covers a fixed number of simple back-and-forth messages. A single agent task can use ten to fifty times the tokens of a single chatbot reply, so a flat £20 a month chatbot plan and a pay-per-token agent workflow aren't directly comparable.

Can I set a hard limit so an agent can't overspend?

Yes. Most API providers, including OpenAI and Anthropic, let you set monthly spend caps and usage alerts, and most agent frameworks let you cap the number of steps or tool calls per task. Both should be switched on before you run anything on live data, not after.

Why does an AI agent's bill vary so much week to week?

Because token use depends on what the agent encounters, not just what it's asked to do. A clean run through simple data costs far less than a run where it hits errors, broken pages, or ambiguous instructions, since every retry and every round of internal checking adds more tokens.

Is it cheaper to build my own agent or buy one off the shelf?

Building your own directly on an API is usually cheaper per task, sometimes by a large margin, because you're not paying a vendor's markup on top of the token cost. The trade-off is you carry the risk of loops and errors yourself, with no support team to catch it, so it suits businesses with someone technical enough to monitor and cap usage.

Where to check the details

• Anthropic's own documentation
• the NIST AI Risk Management Framework