Reduce Your OpenClaw AI Costs by 97%

The Real Cost Numbers

I spent weeks digging through GitHub issues, Discord threads, and real billing data to understand where OpenClaw costs actually come from. The software itself is free and MIT-licensed. The cost is entirely in API calls. Here's what the data shows.

While researching costs across community forums and GitHub discussions, I found users reporting 180 million tokens burned in a single month — roughly $3,600 in API costs for what's supposed to be a "free" tool. And this isn't an isolated case. Here's one example that caught my attention:

This isn't an outlier. Here's what I found across community reports:

90% of users spend under $12/day. But the tail is long. One misconfigured heartbeat can burn $50/day. An automation loop left running overnight can rack up $200-500.

The optimizations below target specific cost drivers. Each one is independent. Start with the one that matches your biggest spending category.

Current Claude API Pricing (Feb 2026)

The gap between Sonnet and Haiku is 3x on input, 3x on output. For tasks that don't need Sonnet's reasoning, you're paying triple for nothing. And DeepSeek V3 is 1/100th the cost of frontier models for simple tasks.

The Hidden 93.5% Token Waste Bug

This is the single biggest cost driver most users don't know about.

There's a proposed fix in the pipeline: a config option to change injection from "always" to "first-message-only." Until that ships, I've been using the session initialization technique below as a workaround.

The community is actively working on this. But right now, the workaround I've been using is straightforward: tell your agent exactly what to load.

Part 1: Session Initialization

One user reported their main session occupied 56-58% of the 400K context window — over 200K tokens consumed before they typed a single word.

The Fix

Add this session initialization rule to your agent's system prompt. It tells the agent exactly what to load and what to skip:

SESSION INITIALIZATION RULE:

On every session start:
1. Load ONLY these files:
   - SOUL.md
   - USER.md
   - IDENTITY.md
   - memory/YYYY-MM-DD.md (if it exists)

2. DO NOT auto-load:
   - MEMORY.md
   - Session history
   - Prior messages
   - Previous tool outputs

3. When user asks about prior context:
   - Use memory_search() on demand
   - Pull only the relevant snippet with memory_get()
   - Don't load the whole file

4. Update memory/YYYY-MM-DD.md at end of session with:
   - What you worked on
   - Decisions made
   - Leads generated
   - Blockers
   - Next steps

Why It Works

Session starts with 8KB instead of 50KB. History loads only when asked. Daily notes become your actual memory. Works with any interface.

Part 2: Model Routing

Out of the box, OpenClaw defaults to Sonnet for everything. Sonnet is excellent. It's also overkill for checking file status, running simple commands, or routine monitoring.

This is the single highest-impact optimization. Users report 50-80% savings from model routing alone.

After testing model routing across multiple setups, the results were consistent: 50-80% cost reduction just by defaulting to Haiku and reserving Sonnet for complex reasoning. Others in the community are seeing the same thing:

The difference between $0.003/1K tokens and $0.0005/1K is 6x. Most tasks simply don't need frontier-level intelligence.

Step 1: Update Your Config

Your OpenClaw config file lives at ~/.openclaw/openclaw.json:

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-haiku-4-5"
      },
      "models": {
        "anthropic/claude-sonnet-4-5": {
          "alias": "sonnet"
        },
        "anthropic/claude-haiku-4-5": {
          "alias": "haiku"
        }
      }
    }
  }
}

This sets Haiku as the default and creates aliases so your prompts can say "use sonnet" or "use haiku" to switch on-demand.

Step 2: Add Routing Rules to System Prompt

MODEL SELECTION RULE:

Default: Always use Haiku
Switch to Sonnet ONLY when:
- Architecture decisions
- Production code review
- Security analysis
- Complex debugging/reasoning
- Strategic multi-project decisions

When in doubt: Try Haiku first.

The Ultra-Cheap Tier

For tasks that barely need intelligence (heartbeats, pings, simple lookups), you can go even cheaper than Haiku:

Part 3: Heartbeat to Ollama

OpenClaw sends periodic heartbeat checks to verify your agent is running. By default, these use your paid API. Running 24/7, that's real money for a health check.

The fix: route heartbeats to a free local LLM using Ollama. This is fully documented in the OpenClaw docs.

Step 1: Install Ollama

# macOS / Linux
curl -fsSL https://ollama.ai/install.sh | sh

# Pull a lightweight model for heartbeats
ollama pull llama3.2:3b

Step 2: Configure OpenClaw for Ollama Heartbeat

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-haiku-4-5"
      },
      "models": {
        "anthropic/claude-sonnet-4-5": { "alias": "sonnet" },
        "anthropic/claude-haiku-4-5": { "alias": "haiku" }
      }
    }
  },
  "heartbeat": {
    "every": "1h",
    "model": "ollama/llama3.2:3b",
    "session": "main",
    "target": "slack",
    "prompt": "Check: Any blockers, opportunities, or progress updates needed?"
  }
}

Step 3: Verify

# Make sure Ollama is running
ollama serve

# In another terminal, test the model
ollama run llama3.2:3b "respond with OK"
# Should respond quickly with "OK"

Part 4: Rate Limits & Budget Controls

Even with model routing and optimized sessions, runaway automation can still burn through tokens. These rate limits act as guardrails.

I've seen multiple reports in GitHub discussions of users waking up to $200-500 bills from agents stuck in repetitive task loops overnight. One common pattern: a failed operation the agent keeps retrying indefinitely. Without rate limits, there's nothing to stop the bleeding.

Add to Your System Prompt

RATE LIMITS:

- 5 seconds minimum between API calls
- 10 seconds between web searches
- Max 5 searches per batch, then 2-minute break
- Batch similar work (one request for 10 leads, not 10 requests)
- If you hit 429 error: STOP, wait 5 minutes, retry

DAILY BUDGET: $5 (warning at 75%)
MONTHLY BUDGET: $200 (warning at 75%)

Workspace File Templates

Keep these lean. Every line costs tokens on every request.

# SOUL.md

## Core Principles
[YOUR AGENT PRINCIPLES HERE]

## Model Selection
Default: Haiku
Switch to Sonnet only for: architecture, security, complex reasoning

## Rate Limits
5s between API calls, 10s between searches, max 5/batch then 2min break

# USER.md

- **Name:** [YOUR NAME]
- **Timezone:** [YOUR TIMEZONE]
- **Mission:** [WHAT YOU'RE BUILDING]

## Success Metrics
- [METRIC 1]
- [METRIC 2]
- [METRIC 3]

Part 5: Prompt Caching

Your system prompt, workspace files, and reference materials get sent with every API call. Prompt caching gives you a 90% discount on repeated content.

How It Actually Works

What to Cache vs Skip

File Structure for Maximum Cache Hits

/workspace/
  ├── SOUL.md                     ← Cache (stable)
  ├── USER.md                     ← Cache (stable)
  ├── TOOLS.md                    ← Cache (stable)
  ├── memory/
  │   ├── MEMORY.md               ← Don't cache (updated frequently)
  │   └── 2026-02-03.md           ← Don't cache (daily notes)
  └── projects/
      └── [PROJECT]/REFERENCE.md  ← Cache (stable docs)

Enable in Config

{
  "agents": {
    "defaults": {
      "cache": {
        "enabled": true,
        "ttl": "5m",
        "priority": "high"
      },
      "models": {
        "anthropic/claude-sonnet-4-5": {
          "alias": "sonnet",
          "cache": true
        },
        "anthropic/claude-haiku-4-5": {
          "alias": "haiku",
          "cache": false
        }
      }
    }
  }
}

Real-World Example: 50 Outreach Drafts/Week

Savings: $70/month on a single workflow.

When NOT to Cache

• Haiku tasks — too cheap to justify caching overhead.
• Frequent prompt changes — cache invalidation costs more than caching saves.
• Small requests (<1KB) — caching overhead eats the discount.
• Development/testing — too many prompt iterations cause cache thrashing.

Part 6: Monitoring & Observability

You can't optimize what you can't measure. These tools show where your tokens actually go.

Verify Your Setup

# Start a session
openclaw shell

# Check current status
session_status

# You should see:
# - Context size: 2-8KB (not 50KB+)
# - Model: Haiku (not Sonnet)
# - Heartbeat: Ollama/local

Cache Performance Check

# Check cache effectiveness
openclaw shell
session_status

# Look for cache metrics:
# Cache hits: 45/50 (90%)
# Cache tokens used: 225KB (vs 250KB without cache)
# Cost savings: $0.22 this session

Troubleshooting

Quick Reference Checklist

The Bottom Line

No complex infrastructure changes needed. After testing all of these across multiple setups, the pattern is clear: smart config, clear rules in your system prompt, a free local LLM for heartbeats, and monitoring to know where your tokens go.

Combined result: from $1,500+/month down to $30-50/month.

The intelligence is in the prompt, not the infrastructure.

Further Reading