Your
OpenClaw Is Burning Money. Here’s Exactly Where It Goes.

Federico Viticci burned through 1.8 million tokens in a month. His
bill hit $3,600. A Reddit user reported spending $200 in a single day
because an automated task got stuck in a loop. A viral Medium post put
the weekly cost at $47 for moderate use. And across every OpenClaw
forum, Discord server, and Reddit thread, the same question keeps coming
up: why is this thing so expensive?

OpenClaw is free. The software costs nothing. You can download it
right now and have a personal AI assistant running in minutes. But the
API calls that power it? Those are very much not free. And if you don’t
understand where the tokens go, you will get a bill that makes you
reconsider your life choices.

Let’s fix that.

The Six Token Vampires

Every token your OpenClaw agent consumes falls into one of six
categories. Understanding them is the difference between a $10/month
setup and a $300/month money pit.

Context accumulation is the big one. It accounts for
40-50% of your total token spend. Every time your agent responds, it
sends the entire conversation history to the API. Message one costs a
few hundred tokens. Message fifty costs tens of thousands — because the
model needs to re-read everything that came before. OpenClaw has
compaction (it summarises old context to keep things manageable), but if
you’re having long, complex conversations, the context window fills up
fast.

Tool output storage eats another 20-30%. When your
agent runs a shell command, reads a file, or fetches a web page, the
output gets stored in the conversation history. A single
ls -la might return a few hundred tokens. A web fetch can
dump thousands of tokens into context. Multiply that across a session
with dozens of tool calls and you’ve got a problem.

The system prompt is a silent killer at 10-15%.
OpenClaw’s system prompt — the instructions that tell your agent who it
is, what tools it has, what files to read — gets sent with every single
API call. If you’ve got a detailed SOUL.md, extensive AGENTS.md, and a
dozen skill files loaded, your system prompt alone could be 5,000-10,000
tokens. Every. Single. Message. Our guide to
SOUL.md is worth reading here, because a bloated personality file
isn’t just bad writing — it’s expensive writing.

Multi-round reasoning takes 10-15%. Complex tasks
require the agent to think, use tools, think again, use more tools, and
iterate. Each round is a fresh API call with the full context. A task
that takes five rounds costs five times what a single-shot response
costs.

Thinking mode is the premium tax. If you’re running
with extended thinking enabled, the model generates internal reasoning
tokens before its visible response. These tokens cost money. On
Anthropic’s Claude, thinking tokens are billed at the same rate as
output tokens. A response that takes 2,000 thinking tokens and 500
visible tokens costs you for all 2,500.

Cron jobs and heartbeats are the silent drain. If
your agent checks email every 30 minutes, runs a weather check four
times a day, and has a heartbeat every 15 minutes — that’s dozens of API
calls running 24/7 whether you’re using your agent or not. Each one
loads the full system prompt, checks context, and generates a response.
Most of the time that response is just “nothing to report.” You’re
paying for an AI to tell you it has nothing to say.

The Model Choice Multiplier

Everything above gets multiplied by your model choice, and this is
where most people’s budgets explode.

Anthropic’s Claude Opus costs roughly $15 per million input tokens
and $75 per million output tokens. Claude Sonnet is $3/$15. Claude Haiku
is $0.80/$4. Google’s Gemini Flash is effectively free for light use.
OpenAI’s GPT-4o sits somewhere in the middle.

If you’re running Claude Opus for everything — every heartbeat, every
cron check, every “what’s the weather?” question — you’re paying luxury
car prices for a bicycle commute. The model that writes your morning
briefing doesn’t need to be the same model that checks if you have new
email.

This is the single biggest lever you have. Our breakdown of free models and our
guide to running OpenClaw
under $5/month both hammer this point, and for good reason.

The Real-World Numbers

Based on community reports and the data available, here’s roughly
what different usage patterns cost:

A light user — daily Q&A, simple tasks, maybe 5-20 messages a day
— burns through 5-20 million tokens monthly. That’s $10-30 on Claude
Sonnet, or under $5 if you’re strategic about model routing.

A medium user — automated workflows, email checking, calendar
management, regular cron jobs — hits 20-50 million tokens. That’s $30-70
on mid-tier models, though it can spike much higher if you’re not
careful.

A heavy user — running OpenClaw as a 24/7 personal assistant with
browser automation, multiple integrations, and constant monitoring — can
blow through 50-200 million tokens. At that point you’re looking at
$70-150+ monthly, and the extreme cases (like Viticci’s $3,600 month)
happen when thinking mode, expensive models, and runaway automation
combine into a perfect storm.

How to Cut Your Bill by 80%

This isn’t theoretical. These are the specific changes that take a
$100/month OpenClaw setup down to $20 or less.

Route cheap tasks to cheap models. Your main
conversation can run on Claude Sonnet or GPT-4o. But sub-agents, cron
jobs, heartbeat checks, and routine automation should run on Gemini
Flash, GPT-4o-mini, or Haiku. OpenClaw supports per-task model
overrides. Use them. There’s zero reason to burn Opus tokens on a
weather check.

Restrict cron jobs to waking hours. If your agent
checks email at 3am and you’re asleep, that’s money burned for nothing.
Schedule cron jobs for 8am-10pm only. Batch checks together — one
heartbeat that covers email, calendar, and weather is cheaper than three
separate cron jobs.

Trim your system prompt. Audit your SOUL.md,
AGENTS.md, and loaded skills. Every line in those files gets sent with
every API call. If your SOUL.md is 2,000 words, consider whether it
needs to be. Shorter personality files aren’t just cleaner — they’re
cheaper.

Use compaction aggressively. OpenClaw compacts
conversation history to manage context size. Make sure it’s working.
Long sessions without compaction mean exponentially growing context,
which means exponentially growing costs.

Go local for the basics. If you’ve got the hardware,
running local models for
routine tasks eliminates token costs entirely. Ollama with a capable
open-source model handles basic Q&A, file operations, and simple
automation at zero marginal cost. Keep the cloud models for tasks that
actually need them.

Monitor before you optimise. You can’t fix what you
can’t see. Use openclaw status --usage to check your token
consumption. Track which tasks burn the most tokens. The answer might
surprise you — it’s often not the complex tasks, but the high-frequency
routine ones that quietly drain your budget.

Kill runaway loops. Set timeouts on automated tasks.
An agent stuck in a retry loop can burn through hundreds of dollars in
hours. Guardrails aren’t optional when you’re paying per token.

The Uncomfortable Truth

OpenClaw’s cost problem isn’t really OpenClaw’s fault. It’s the
economics of large language model APIs in 2026. Every AI assistant that
relies on cloud APIs has this problem. OpenClaw just makes it more
visible because it’s transparent about what it does — you can see every
API call, every token count, every dollar spent.

The good news is that model prices are dropping fast. What cost
$15/million tokens a year ago costs $3 today. Gemini Flash gives you a
million tokens for pocket change. Local models are getting capable
enough to handle 80% of daily tasks. The trajectory points toward a
world where running a full AI assistant costs less than a Netflix
subscription.

We’re not there yet. But if you’re smart about model routing,
ruthless about trimming unnecessary automation, and willing to use cheap
models for cheap tasks, you can run a genuinely useful OpenClaw setup
for the price of a couple of coffees a month.

For the full setup guide on doing exactly that, check our complete cost breakdown and the
step-by-step for running
under $5/month. Your wallet will thank you.