What an AI agent costs per month, worked out

An AI agent costs its runs a month times the cost of one run, and a run costs the tokens of every turn, because each turn sends the whole conversation again. On Claude Sonnet 5.5 with prompt caching, a 6-turn support conversation comes to about 4 cents, so 5,000 a month cost about $192; a 20-turn research report about 46 cents; a 150-turn coding session about $7.

The figures below are worked out turn by turn from the makers' prices, with the same sums as our agent cost simulator. Change the agent and its settings, and the monthly bill on each current Claude model follows.

Try itWhat an agent costs a month

Agent

·A support bot·A research agent·A long coding session·Anthropic's 40-turn task

Prompt cache

OnOff

Turns in a run

6turns

each one a call to the model

110100

Runs a month

5Kruns

conversations, reports or sessions

101K100K

Tokens sent on each turn

New, full priceRead from the cache

1. Claude Opus 5.57.7 cents a run$385 a month
2. Claude Sonnet 5.53.8 cents a run$192 a month
3. Claude Haiku 5.50.21 cents a run$10.33 a month

Why a run costs more than its prompt

A chat answer is 1 call to the model. An agent is many: it calls a tool, reads the result, decides what to do next and calls the model again. The model keeps no memory between calls, so every call sends the whole conversation so far: the system prompt, the tool descriptions, the task, and everything each earlier turn wrote and got back.

So the prompt grows on every turn, and the run pays for all of them. Anthropic's example is a task that starts at 20,000 tokens and grows to 120,000 over 40 turns. The conversation is never larger than 120,000 tokens, but the 40 calls send about 2.8 million input tokens between them. With no cache that is $11.20 of input on Claude Opus 5.5.

2 things follow:

• Prompt caching matters more for agents than for anything else. All but the newest part of each turn's prompt was sent on the turn before. With 90% read from the cache, Anthropic's 2.8 million tokens cost about $1.62.
• Cost grows faster than the number of turns. Each extra turn is the most expensive one so far. The research agent below at 20 turns costs 46 cents a report with the cache; at 40 turns, $1.13. With no cache, $2.37 becomes $9.06.

3 agents, worked out

Each run's tokens and the price of a run on Claude Sonnet 5.5, with the cache on and off:

A month of each, cached:

• Support: 5,000 conversations cost $192 on Sonnet 5.5, $385 on Opus 5.5 and about $11 on Claude Haiku 5.5.
• Research: 20 reports a working day, 440 a month, cost $200 on Sonnet 5.5.
• Coding: 2 long sessions a working day cost about $320 a month per developer on Sonnet 5.5 and $640 on Opus 5.5. For a measured comparison, Anthropic puts Claude Code at about $13 per developer per active day across enterprise deployments, and under $30 for 90% of users.

The figures for each turn's writing and tool results are our own estimates; your agent's logs give the real ones. Every response reports its input, output and cache tokens, and those are what to multiply.

One trap on small models: Claude Haiku 5.5 charges 5 times its usual prices for any request whose prompt is over 100,000 tokens, cache reads included. The coding session passes that line on turn 21 of 150, so on Haiku it costs $2.83 with the cache, about 2 fifths of the Sonnet price instead of a twentieth.

Against a salary

The 2025 US median pay for a customer service representative was $44,770 a year, about $3,731 a month before benefits. The support bot's tokens for 5,000 conversations cost about $192 on Sonnet 5.5. That comparison is real but incomplete, in 3 ways:

• An agent handles the share of conversations it can, and hands the rest to a person. The saving is in that share, not the whole desk.
• Someone has to build it, write its tests, read its failures and keep it current. That is engineering time, every month.
• A wrong refund or a wrong answer has a cost the token bill does not show. How to evaluate an AI agent covers how to measure that before it ships.

What the token bill leaves out

• Server tools. On the Claude API web search costs $10 per 1,000 searches, on top of the tokens its results add. Code execution is free for 1,550 hours a month per organization, then $0.05 an hour per container.
• A hosted agent. Claude Managed Agents bills the same token prices plus $0.08 for each hour a session is running.
• Your own servers. The service that runs the loop, a vector database if the agent searches documents (the vector database size calculator sizes one), and the logs you keep of every run.

How to bring it down

1. Cache. Keep the system prompt and tool list identical between calls and turn caching on. For agents it is usually the largest single cut; how to reduce LLM API costs goes through the steps in order.
2. Keep runs short. Since each turn costs more than the last, trimming old tool results or summarising a long conversation saves on every later turn. Anthropic estimates that compacting at 150,000 tokens pays for itself within about 10 turns.
3. Write less. Output, thinking included, costs 100 times a cache read on Opus 5.5. A lower effort setting on simple steps cuts it.
4. Test a smaller model, remembering Haiku 5.5's long-prompt prices for long runs.

To price your own agent, enter its first prompt, what each turn adds and its turns in the agent cost simulator, which compares every model with a listed price. What an AI agent isAgentA model that uses tools in a loop, deciding each next step from what the last one returned, until the task is done. explains the loop of turns itself.