Reduce Your OpenClaw Costs by 97%: Token Optimization Guide

If you’ve been running OpenClaw and watching your API bills climb, you’re not alone. This guide covers five optimizations that work together to slash costs: session initialization, model routing, local heartbeat, rate limits, and prompt caching.

1. Session Initialization

The fix: add a session initialization rule to your agent’s instructions. Tell it exactly what to load — and what NOT to load — at session start.

# Session Initialization Rule (add to AGENTS.md)

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
   - Don't load the whole file

4. Update memory/YYYY-MM-DD.md at end of session:
   - What you worked on
   - Decisions made
   - Blockers and next steps

2. Model Routing

Out of the box, OpenClaw defaults to Claude Sonnet for everything. While Sonnet is excellent, it’s overkill for checking file status, running simple commands, or routine monitoring. Haiku handles these perfectly at a fraction of the cost. For a full breakdown of which models to use when, see our model selection guide.

// ~/.openclaw/openclaw.json
{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-haiku-4-5"
      }
    }
  }
}

Then add routing rules to your agent instructions:

# Model Selection Rule (add to AGENTS.md)

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.

3. Route Heartbeat to Haiku

OpenClaw sends periodic heartbeat checks to verify your agent is alive. By default, these use your primary model — which adds up fast when running 24/7. Route them to Haiku 4.5, the cheapest Claude model, and reduce the frequency to once per hour.

// Add to ~/.openclaw/openclaw.json
{
  "heartbeat": {
    "every": "1h",
    "model": "anthropic/claude-haiku-4-5",
    "session": "main",
    "prompt": "Check: Any blockers or progress updates?"
  }
}

4. Rate Limits & Budget Controls

Even with model routing and optimized sessions, runaway automation can burn through tokens. These guardrails prevent accidental cost explosions.

# Rate Limits (add to AGENTS.md)

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 for 1 lead each)
- If you hit 429 error: STOP, wait 5 minutes, retry

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

5. Prompt Caching

Your system prompt, workspace files, and reference materials get sent to the API with every single message. Prompt caching (available on Claude 3.5 Sonnet and newer) charges only 10% for cached tokens on re-use. For static content you send repeatedly, this cuts costs by 90%.

What to cache vs. not

// Enable caching in ~/.openclaw/openclaw.json
{
  "agents": {
    "defaults": {
      "cache": {
        "enabled": true,
        "ttl": "5m",
        "priority": "high"
      }
    }
  }
}

Maximize cache hits

• +Batch requests: Make multiple API calls within 5-minute windows to reuse cached prompts
• +Keep prompts stable: Don't update SOUL.md mid-session — changes invalidate cache
• +Separate stable from dynamic: Reference docs (cached) vs. daily notes (uncached) in different files
• +Target >80% hit rate: Monitor with session_status and adjust if cache misses are high

Combined Impact

Each optimization targets a different cost driver. Together, they compound:

Verifying Your Setup

# Start a session and check status
openclaw shell
session_status

# You should see:
# - Context size: 2-8KB (not 50KB+)
# - Model: Haiku (not Sonnet)
# - Heartbeat: Haiku (not Sonnet)
# - Cache hit rate: >80%

Troubleshooting

• +Context still large: Check session initialization rules are in your AGENTS.md
• +Still using Sonnet: Verify openclaw.json syntax and model.primary path
• +Heartbeat errors: Verify heartbeat.model is set to anthropic/claude-haiku-4-5 in openclaw.json
• +Costs haven't dropped: Check that your system prompt is actually being loaded

The Bottom Line

Five optimizations. Five minutes each. Combined result: from $1,500+/month to under $50. That’s money you can reinvest in actually building things.

These savings are especially impactful for high-volume workflows like automated cold outreach or AI-generated content at scale.

On founders.sh, we apply these optimizations by default on every agent container we deploy. You get model routing, session management, and cost controls out of the box — without touching a config file.