💰 Cost Optimization

📊 Where does the money go?

🏗️ Model tiering strategy

The biggest cost lever: use expensive models only when needed.

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-sonnet-4-5",
        "fallbacks": ["openai/gpt-4.1-mini"]
      },
      "heartbeat": {
        "model": "google/gemini-2.5-flash-lite"
      }
    }
  }
}

💓 Cheap heartbeat model

Heartbeat runs every 15-30 minutes. Using Sonnet here wastes money on "check if there's a new email" tasks:

{
  "agents": {
    "defaults": {
      "heartbeat": {
        "every": "30m",
        "model": "google/gemini-2.5-flash-lite",
        "activeHours": {
          "start": "08:00",
          "end": "22:00",
          "timezone": "Europe/Bucharest"
        }
      }
    }
  }
}

• Flash-Lite for heartbeat: ~$0.50/mo vs ~$3/mo with Sonnet
• Active hours: No heartbeats while you sleep = 33% savings
• Increase interval: "every": "1h" if 30 min is too frequent

⏰ Cron job optimization

Cron jobs are the second biggest cost driver. Key strategies:

Use cheaper models per cron job

openclaw cron add \
  --name "Dependency audit" \
  --cron "0 9 * * 1" \
  --model "google/gemini-2.5-flash-lite" \
  --message "Run npm audit and npm outdated"

Use --session isolated

Isolated sessions prevent cron jobs from inflating your main session's context (and cost):

openclaw cron add \
  --name "Morning briefing" \
  --cron "0 7 * * *" \
  --session isolated \
  --message "Send my morning briefing"

Set session retention

{
  "cron": {
    "sessionRetention": "24h",
    "runLog": {
      "maxBytes": "2mb",
      "keepLines": 2000
    }
  }
}

Sessions from completed cron runs are pruned after 24h, preventing bloat.

📏 Context & token management

• Daily session reset: Prevents conversations from growing indefinitely — config example:

{
  "session": {
    "reset": {
      "mode": "daily",
      "atHour": 4,
      "idleMinutes": 120
    }
  }
}

• Auto-compaction: OpenClaw automatically compacts context when it overflows — but better to reset before that happens
• Idle timeout: idleMinutes: 120 resets the session after 2 hours of inactivity
• Shorter SOUL.md: Every word in SOUL.md is sent with every message. Trim unnecessary instructions.

📦 Skills optimization

Every enabled skill gets injected into the system prompt on every API call — even when irrelevant. 20 custom skills ≈ ~32,000 extra tokens per message, adding $30–100/month.

Quick wins

• Trim skills.allowBundled: The default loads all 53 bundled skills. Keep only 3–5 essentials you use on every session.
• Archive custom skills: Move specialist skills to ~/.openclaw/skills-archive/ and load them on demand via a Skill-Router or IDENTITY.md Relay.

{
  "skills": {
    "allowBundled": ["github", "healthcheck", "session-logs"]
  }
}

Result: Production setups have cut skill token costs by up to 89% (e.g. ~125k → ~19k tokens per call). For the full pattern — Skill-Router vs IDENTITY.md Relay, implementation steps, and real-world benchmarks — see How to Load Skills On Demand.

🖥️ Local models (zero cost)

Run models locally with Ollama for zero API cost:

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-sonnet-4-5",
        "fallbacks": ["ollama/qwen2.5:7b"]
      },
      "heartbeat": {
        "model": "ollama/qwen2.5:7b"
      }
    }
  }
}

Local fallback serves two purposes: saves money on simple tasks, and keeps working when cloud APIs go down.

🚨 Budget limits & alerts

OpenRouter spending limits

If using OpenRouter as your provider, set a monthly cap on your API key:

# In OpenRouter dashboard:
# Settings → API Keys → Set monthly limit: $30

This is critical for Telegram bots that are always online — a runaway conversation loop can burn through credits fast.

Per-model rate limiting

{
  "agents": {
    "defaults": {
      "model": {
        "rateLimit": {
          "maxRequestsPerMinute": 10,
          "maxTokensPerDay": 500000
        }
      }
    }
  }
}

📋 Cost profiles

📈 Usage tracking

# Check current session cost
/status

# Check model usage across all sessions
openclaw status --deep

# View cron job costs
openclaw cron runs --id <job-id>

# Check OpenRouter spend
# Visit openrouter.ai/activity