Guide 09 · Cost · 3 min read

The cost of running agents

In brief

Agents don't charge per answer — they charge per step. Every step re-sends everything the agent has read and done so far. So an agent task costs several times more than a chat: roughly 4× for one agent, and 15× for a team of agents. And the model bill is only the part you can see.

A chat is one call to the AI model. An agent calls the model again and again — and each call re-reads everything so far. Once you see that, the bill makes sense.

Fig. 09 — agents bill by the loop, not by the answer

Start with how the meter works. AI models charge for tokens — small chunks of text, counted going in and coming out. An agent works in steps. Each step re-sends everything so far: its instructions, the conversation, and every tool result. Ten steps means paying for the early text ten times over. That's why Anthropic's real-world numbers land where they do: one agent uses roughly 4× the tokens of a plain chat, and a team of agents roughly 15× [1].

So how do you cut the bill? Prompt caching helps most. Caching means the model saves the text it keeps re-reading, and reads the saved copy back at a tenth of the normal input price. That matters a lot when the same opening text is re-sent on every step [2]. Two more levers: send the simple, routine steps to smaller and cheaper models, and turn steps you already know into a fixed workflow instead of a free loop. Both make each step shorter or cheaper — the old rule of picking the simplest thing that works [3].

Count the cost per finished task, not per token.

What costs don't show up on the bill?

The token bill is the easy part — you can screenshot it. The rest is quieter. Someone has to check the agent's work. Every tool the agent uses becomes something you now have to keep working. And when a failed run has already changed things in the real world, the cleanup costs more than the tokens did. The honest number folds all of this in — cost per successfully finished task, failed runs included. That's also why your tests should track cost and speed next to the success rate.

Gas meter with a mechanical counting dial
Plate 09 — the meter runs whether or not you read it Photo — Arthur Lambillotte, Unsplash

Sources

  1. Anthropic — How we built our multi-agent research system, 13 Jun 2025 (token multiples)
  2. Claude docs — Prompt caching (platform.claude.com; cache reads at 0.1× base input price)
  3. Anthropic — Building effective agents, 19 Dec 2024 (simplest solution first)

AiAgentNook editorial · corrections are noted here, not silently edited.

Frequently asked questions

How much does running an AI agent cost?

It depends on how many steps the agent takes, not how many answers it gives. Each step re-sends everything so far. Anthropic's real-world data puts one agent at roughly 4× the tokens of a chat, and a team of agents at roughly 15×. Judge cost per finished task, and count the failed runs too.

How can I reduce AI agent costs?

Four levers. Use prompt caching, which lets the model re-read saved text at about a tenth of the normal price. Send simple steps to smaller, cheaper models. Turn steps you already know into a fixed workflow instead of a free loop. And set step and spending limits, so failed runs stop early instead of burning money.

Why do AI agents use so many tokens?

Tokens are the small chunks of text a model charges for. On every step, an agent re-sends everything so far — its instructions, the history, and all earlier tool results — plus its new output. Ten tool calls means paying for the early text ten times. Long histories and wordy tool results make it worse.