5 OpenClaw Mistakes Costing You Money Right Now | OpenClaw DC

Five settings are silently draining your OpenClaw budget. The heartbeat background process alone costs $50-150/month. Using API tokens instead of your existing subscription wastes $10-30/day. Long conversation sessions multiply your token cost with every message. Here are the five mistakes and how to fix each one in under 10 minutes.

TL;DR: You are probably making at least three of these five mistakes. (1) Paying per token when your ChatGPT/Codex subscription covers it. (2) Letting the heartbeat send 100K tokens every 30 minutes for free. (3) Using expensive models for simple tasks. (4) Letting conversation history balloon your costs. (5) Running your cloud server 24/7. Fix all five and a $300/month bill drops to $40-80/month. For teams, $36K/year becomes $5-10K/year.

Mistake 1: Paying Per Token When You Already Have a Subscription

This is the most common mistake and the easiest to fix. Most OpenClaw users set up API keys and start paying per token without realizing they already have a ChatGPT Plus or Codex subscription that covers the same models. That subscription costs a flat monthly fee. API tokens cost $10-30/day depending on usage.

The math is simple. If you are paying $20/month for ChatGPT Plus and also spending $15/day on API tokens, you are throwing away $450/month for access you already have.

How to fix it:

1. Check if you have an active ChatGPT Plus, Team, or Codex subscription
2. Configure OpenClaw to route through your subscription instead of raw API keys
3. Update your provider config:

# config.yaml
providers:
  openai:
    mode: subscription
    plan: plus  # or team, or codex

Savings: $10-30/day ($300-900/month)

Mistake 2: The Heartbeat Is Bleeding You Dry

This one surprised most people in the video comments. OpenClaw runs a background process called the heartbeat that checks in on your tasks periodically. By default, it fires every 30 minutes. That sounds harmless until you realize what it actually does.

Every heartbeat sends your entire context window to the model. That means 100K+ tokens going out every 30 minutes, 24 hours a day, 7 days a week. You are paying for a full model inference call while you sleep, eat dinner, and sit in meetings.

At typical token rates, the heartbeat alone costs $50-150/month doing absolutely nothing useful most of the time.

How to fix it (four changes, pick any combination):

Option A: Turn it off entirely. If you do not need background task monitoring, disable it.

# config.yaml
heartbeat:
  enabled: false

Option B: Use a cheap model. The heartbeat does not need Claude Opus or GPT-4 to check a task list. Route it to a model that costs pennies.

heartbeat:
  model: "haiku"
  slim_context: true

Option C: Business hours only. No reason to run heartbeats at 3 AM.

heartbeat:
  schedule:
    start: "08:00"
    end: "18:00"
    timezone: "America/New_York"

Option D: Isolated sessions. This is the biggest win. Instead of sending your full 100K-token context, use an isolated heartbeat session that only sends the task queue. This drops the per-heartbeat cost from 100K tokens down to 2-5K tokens.

heartbeat:
  isolated_session: true
  include_history: false

Combining options B, C, and D together takes the heartbeat from $50-150/month down to under $2/month.

Savings: $48-148/month

Mistake 3: Using Expensive Models for Simple Tasks

This one came up in the video as “paying a brain surgeon to take your temperature.” It is a perfect analogy.

Most of what OpenClaw does during a typical session is simple: classifying a message, formatting text, acknowledging a task, parsing a date. These tasks do not need a $25-per-million-token model. A model that costs $0.25 per million tokens produces identical results for these operations.

Yet the default configuration sends everything to the same model regardless of complexity. A “summarize this email” request and a “write a complex multi-step automation” request both hit your most expensive model.

How to fix it:

# config.yaml
model_routing:
  enabled: true
  default_model: "haiku"      # cheap, handles 80% of tasks
  complex_model: "sonnet"     # expensive, reserved for hard work
  complexity_threshold: 0.7

With this configuration, roughly 80% of requests go to the cheap model and only complex, multi-step reasoning tasks escalate to the premium model. Quality stays the same because the cheap model handles simple tasks perfectly well.

Savings: 80-90% on model costs

Mistake 4: Never Starting Fresh Sessions

This is the silent cost multiplier. Every time you send a message in OpenClaw, the entire conversation history gets sent along with it. After a few turns, you are carrying roughly 13K tokens of background information before your actual question even reaches the model.

After ten turns, you are sending a small novel. After thirty turns, you are paying 10-15x what the same work would cost in a fresh session. One user in the video traced a single debugging session that consumed more tokens than their entire first week of usage.

The fix is simple: use /new when you switch tasks.

When to start a fresh session:

• Switching from one project to another
• Moving from debugging to writing
• Starting a new type of task
• Anytime the conversation is over 10 turns and the topic has changed

How to make it a habit:

# In your OpenClaw session, just type:
/new

That is it. No configuration needed. Just build the muscle memory to type /new when you pivot to a different task. Your context drops from thousands of accumulated tokens back to your base system prompt.

Savings: 50-70% on token costs

Mistake 5: Running Your Cloud Server 24/7

If you are self-hosting OpenClaw on a cloud server (AWS, GCP, Azure, or a VPS), the compute meter runs all day whether you are using it or not. Most developers work 8-10 hours a day. That means 14-16 hours of idle server time burning money.

How to fix it:

Schedule your server to start when you start working and stop when you stop.

# AWS example: start at 8 AM, stop at 6 PM Eastern
aws events put-rule \
  --name "openclaw-start" \
  --schedule-expression "cron(0 13 ? * MON-FRI *)"

aws events put-rule \
  --name "openclaw-stop" \
  --schedule-expression "cron(0 23 ? * MON-FRI *)"

# Or use a simple cron job on any VPS
# crontab -e
0 8 * * 1-5  systemctl start openclaw
0 18 * * 1-5 systemctl stop openclaw

If you only run during business hours on weekdays, you cut server costs by roughly 60-70%.

Savings: 60-70% on compute costs

Team Cost Comparison

These mistakes compound across a team. Here is what the numbers look like for a 10-developer team:

Even using more conservative estimates, the shift from $36K/year unoptimized to $5-10K/year optimized is realistic for most teams.

The Playbook: Set This Up for Any Team

Here is the checklist. Do these in order.

• Audit existing subscriptions (ChatGPT Plus, Codex, Team plans) and route through them
• Set heartbeat to isolated session + cheap model + business hours only
• Enable model routing with a cheap default and expensive fallback
• Train the team to use /new when switching tasks (add it to onboarding docs)
• Schedule cloud servers to start/stop with business hours
• Run /usage weekly to track spend and catch regressions
• Set per-user spending alerts at 50% and 80% of budget

Related Guides

• OpenClaw Costs $600/Month? Here’s How to Get It Under $20
• OpenClaw Token Usage Explained
• The Complete OpenClaw Costs Guide
• How to Cut Your OpenClaw API Costs by 80%

More OpenClaw Videos

We publish weekly videos on OpenClaw setup, optimization, and real-world use cases. Check out the full playlist on our YouTube channel for more walkthroughs, including deep dives on model routing, self-hosting, and automation builds.

Need help setting this up for your team? We configure OpenClaw environments for teams across the DC metro area and remotely.

Book a Call