Cubicles
Cubicles

How much does it cost to run an AI agent? A realistic monthly budget

OpenClaw, Hermes and most coding agents are free to download. Running them isn't. Here's where the money goes, how to estimate your own bill, and how to keep it from surprising you.

Blog ·

Ask how much an AI agent costs and you'll get answers from “nothing, it's open source” to “I spent $300 in a weekend”. Both can be true. The software is usually free. What you pay for is somewhere for it to run and the model it thinks with.

The two bills

  1. The computer. Somewhere the agent runs all day: your laptop, a Mac mini, a server or a cloud computer. This cost is predictable.
  2. The model. Every step the agent takes is a call to an AI model, billed by your provider per token. This cost depends entirely on how you use it.

Some agents can also sign in with a consumer subscription instead of an API key when you set them up yourself. Check the agent's and the provider's terms for always-on use, since subscriptions usually come with usage limits.

Bill 1: the computer

Where it runsTypical costWhat you get
Your laptop$0Fine for testing, stops when it sleeps, shares your files
Mac miniAbout $599 upfrontAlways on, a desktop, on your home network
Headless VPS$40+ a month for 8 GBAlways on, no desktop for apps
Developer sandbox≈ $240+ a month if left onDesktop, but sessions expire
CubiclesFirst one free, then $24 a monthAlways on, a desktop, off your network

Prices are estimates of public pricing as of November 2026. For the full trade-offs, see where should your AI agent run?

Bill 2: the model

This is the bill that varies. You pay per token, roughly per word, for everything the model reads and writes. An agent reads far more than it writes, because on every step it re-reads:

  • Its instructions and memory. Standing instructions, memory files and skill descriptions go in with every call.
  • The conversation so far. Each step includes the steps before it, so long tasks get more expensive per step as they go.
  • What it just saw. A web page's text, a command's output or a screenshot. Screenshots are images, and images cost more than text.

On top of that, a few habits multiply the bill:

  • Frequent check-ins. An agent that wakes every five minutes to check something makes 288 calls a day before doing any real work.
  • Loops and retries. A step that keeps failing, retried over and over, is the classic overnight bill.
  • The biggest model for everything. Top-tier models cost many times more than smaller ones that handle routine work fine.

How to estimate your own bill

There's no universal number, because a daily inbox sweep and an all-day research agent are different jobs. The reliable way to estimate is to measure:

  1. Run one typical task, start to finish.
  2. Look at your provider's usage page to see the tokens it used.
  3. Multiply by how often the task runs in a month.
  4. Multiply by your model's price per million tokens, and add 30% for retries and the unexpected.

Illustrative example: a morning inbox sweep uses 200,000 tokens and runs every day, so about 6 million tokens a month. At a made-up blended rate of $3 per million tokens, that's about $18 a month. Use your own model's prices: they vary a lot between models and change often.

Ten ways to cut model spend

  1. Set a hard monthly limit with your provider before anything else. It's the one setting that caps the downside.
  2. Use a smaller model for routine work. Sorting email and checking a page for changes rarely needs the top model. See choosing a model for your agent.
  3. Compare models through one key. A router like OpenRouter lets you try several models on the same task and keep the cheapest that works.
  4. Narrow the job. A focused agent with short instructions reads fewer tokens on every step.
  5. Check less often. Hourly instead of every five minutes is a twelvefold saving on checks.
  6. Keep memory tidy. Prune old notes and unused skills, since they're re-read on every call.
  7. Prefer text to screenshots. Where a page's text or a command's output will do, it's cheaper than an image.
  8. Cap steps and retries. Tell the agent to stop and ask after a few failed attempts rather than trying forever.
  9. Use prompt caching where offered. Many providers charge less for instructions repeated between calls.
  10. One key per agent. Then you can see which agent is expensive, and fix that one.

Three example setups

SetupComputerModelHow to start
One personal assistantOne computer, free on CubiclesA smaller model for routine, a bigger one on requestSet a modest monthly limit and review after a week
Assistant plus research agentTwo computers: free plus $24 a monthA different key per agentWatch which job uses more, then tune it
A coding agent overnightOne computer, more memory for big buildsA strong coding modelCap steps per task and limit spend per night

The computer is the predictable part. Put your attention on the model: measure a week, set limits, and move routine work to smaller models.

What Cubicles charges

On Cubicles, you pay for the computer only. Your first computer is free, with no card needed. More are $24 a month each on the Standard plan, with 4 vCPU, 8 GB of RAM and 64 GB of NVMe storage, and no data caps. Model usage goes straight to your provider through your own API key, so Cubicles never marks it up. Create your free computer.

Frequently asked questions

How much does OpenClaw cost?

OpenClaw itself is free and open source. You pay for the computer it runs on and for model usage with your AI provider. Model usage depends on how much it does and which model you choose.

Why is my AI agent using so many tokens?

Agents re-read their instructions, memory and the conversation so far on every step, and screenshots are expensive. Frequent check-ins, retries and long-running tasks multiply that. Shorter instructions, fewer check-ins and a cap on retries help most.

Can I set a spending limit for my AI agent?

Yes, with your model provider. Most let you set a monthly budget or limit per API key. Give each agent its own key and set a limit on each.

Is it cheaper to run a model locally?

Sometimes, if you already own powerful hardware. Local models need a lot of memory to run well, and smaller ones are often weaker at the tool use agents depend on. For most people, a hosted model with a spending limit is cheaper overall.

Does Cubicles charge for AI model usage?

No. You bring your own API key, and usage is billed by your provider directly. Cubicles charges only for computers: your first is free and more are from $24 a month.

Give your agent a computer