# Cost Optimization: Keep API Bills Under Control Canonical: https://clawdocx.com/docs/cost-optimization Author: Sam Okafor Difficulty: INTERMEDIATE Published: 2026-02-24 Updated: 2026-09-19 > Practical strategies for reducing OpenClaw API costs: what costs money, model routing, active hours, usage monitoring, and spending limits. ## Understanding OpenClaw Costs OpenClaw itself is free and open-source. The costs come from the LLM API providers, Anthropic, OpenAI, Google, that power your agent's intelligence. Every time your agent processes a message, runs a heartbeat cycle, or executes a cron job, it sends tokens to the API and you pay for those tokens. > **Note** > > The number one complaint from new OpenClaw users is unexpected API bills. A carelessly configured agent with Claude Opus 4.6 can easily cost $200-500/month. With proper optimization, you can run a fully functional agent for $10-30/month. ### What Costs Money | Activity | Frequency | Typical Cost per Event | Monthly Impact | |----------|-----------|----------------------|----------------| | **Heartbeat cycles** | Every 10-15 min | $0.002-0.05 | $5-200 | | **User conversations** | Varies | $0.01-0.10 per message | $5-50 | | **Cron jobs** | Scheduled | $0.01-0.05 per execution | $2-30 | | **Skill execution** | On demand | $0.01-0.20 per invocation | $5-50 | | **Context window** | Every request | Included in above | n/a | The heartbeat is typically the largest expense because it runs continuously, even when you are not interacting with the agent. ### Cost by Model Understanding per-token pricing helps you make informed decisions: | Model | Input (per 1M tokens) | Output (per 1M tokens) | Relative Cost | |-------|----------------------|------------------------|---------------| | Claude Opus 4.6 | $15.00 | $75.00 | 15x baseline | | Claude Sonnet 4.6 | $3.00 | $15.00 | 3x baseline | | Claude Haiku 4 | $1.00 | $5.00 | **Baseline** | | GPT-5.2 | $3.00 | $12.00 | 3x baseline | | Gemini 3.1 Pro | $2.50 | $10.00 | 2.5x baseline | | Gemini 3 Flash | $0.10 | $0.40 | 0.1x baseline | > **Tip** > > Claude Haiku 4 and Gemini 3 Flash provide excellent capability for routine tasks at a fraction of the cost. Gemini 3 Flash has a generous free tier. Reserve expensive models like Opus 4.6 and o3 for tasks that truly need them. --- ## Strategy 1: Model Routing The single most effective cost optimization is using different models for different tasks. There is no reason to use Claude Opus 4.6 for checking whether a reminder is due. ### Recommended Model Routing Configuration ```json { "model_routing": { "default": "claude-sonnet-4-6-20260217", "heartbeat": "claude-haiku-4-20260210", "quick_response": "claude-haiku-4-20260210", "complex_reasoning": "claude-sonnet-4-6-20260217", "code_generation": "claude-sonnet-4-6-20260217", "summarization": "claude-haiku-4-20260210", "creative_writing": "claude-sonnet-4-6-20260217" } } ``` ### Cost Impact of Model Routing | Configuration | Estimated Monthly Cost | |--------------|----------------------| | Everything on Opus 4.6 | $200-500 | | Everything on Sonnet 4.6 | $40-100 | | Sonnet 4.6 default + Haiku 4 heartbeat | $15-50 | | Haiku 4 default + Sonnet 4.6 for complex tasks | $10-30 | | Gemini 3 Flash heartbeat + Haiku 4 default | $5-15 | ### Automatic Routing with Smart Escalation OpenClaw can automatically determine when a cheaper model is insufficient and escalate: ```json { "model_routing": { "default": "claude-haiku-4-20260210", "smart_escalation": { "enabled": true, "escalation_model": "claude-sonnet-4-6-20260217", "escalation_triggers": [ "complex_task_detected", "multi_step_reasoning", "code_generation", "long_form_writing", "user_requested" ], "max_escalations_per_hour": 10 } } } ``` With smart escalation, Haiku handles 80-90% of interactions, and Sonnet only activates when needed. This can cut costs by 60-70% compared to using Sonnet for everything. --- ## Strategy 2: Optimize the Heartbeat The heartbeat is your biggest cost lever. See the [Heartbeat & Cron Guide](/docs/heartbeat-config) for detailed configuration. ### Quick Wins **Switch heartbeat to Haiku:** ```json { "heartbeat": { "model": "claude-haiku-4-20260210" } } ``` Savings: 80-95% of heartbeat costs. **Increase the interval:** ```json { "heartbeat": { "interval_minutes": 30 } } ``` Savings: 50% (compared to 15-minute interval). **Set active hours:** ```json { "heartbeat": { "active_hours": { "start": "08:00", "end": "22:00", "timezone": "America/New_York" }, "interval_minutes": 15, "sleep_interval_minutes": 120 } } ``` Savings: 30-40% by reducing overnight cycles. **Combine all three optimizations:** Using Haiku, 30-minute intervals during active hours (8 AM - 10 PM), and 2-hour intervals overnight reduces heartbeat costs from ~$200/month (Opus, 10-min) to ~$3/month. --- ## Strategy 3: Limit Context Window Size Every request includes conversation history (context). Longer context means more input tokens and higher costs. Configure maximum context to control this: ```json { "context": { "max_history_messages": 20, "max_context_tokens": 4096, "summarize_old_context": true, "summary_model": "claude-haiku-4-20260210" } } ``` - **`max_history_messages`**: Keep only the last N messages in context - **`max_context_tokens`**: Hard cap on context size in tokens - **`summarize_old_context`**: When true, old messages are summarized rather than included verbatim - **`summary_model`**: Use a cheap model for summarization > **Note** > > Context summarization is particularly effective for long conversations. Instead of sending 50 messages of history (expensive), OpenClaw summarizes them into a paragraph (cheap) and includes only recent messages verbatim. --- ## Strategy 4: Set Spending Limits ### Provider-Level Limits Set spending limits directly on each provider's dashboard: - **Anthropic**: Console > Settings > Limits > Monthly spending limit - **OpenAI**: Platform > Settings > Limits > Usage limit - **Google**: Cloud Console > Billing > Budgets & alerts ### OpenClaw-Level Limits Configure spending awareness in OpenClaw: ```json { "cost_management": { "daily_budget": 5.00, "monthly_budget": 50.00, "alert_at_percentage": 80, "action_on_limit": "downgrade_model", "downgrade_to": "claude-haiku-4-20260210", "notification_channel": "telegram" } } ``` When the daily or monthly budget is approaching: - At 80%: Send an alert to your notification channel - At 100%: Downgrade to the cheapest model (Haiku) instead of stopping entirely This ensures your agent keeps running but stops using expensive models when the budget is exhausted. --- ## Strategy 5: Optimize Cron Jobs Each cron job is an API call. Review your cron jobs and optimize: ### Use Appropriate Models | Cron Job Type | Recommended Model | |---------------|-------------------| | Email check | Haiku 4 | | Calendar briefing | Haiku 4 | | News summarization | Sonnet 4.6 | | Weekly review | Sonnet 4.6 | | Simple monitoring | Haiku 4 | | Complex analysis | Sonnet 4.6 (sparingly) | ### Reduce Frequency Do you really need to check email every 15 minutes? Every hour is usually sufficient for most users. ### Consolidate Jobs Instead of three separate morning cron jobs (email, calendar, weather), combine them into one: ``` Schedule: 0 8 * * * Task: "Morning briefing: Check my unread emails, today's calendar, and the weather forecast. Give me everything in one summary." ``` One API call instead of three. ### Pause Unnecessary Jobs ```bash openclaw cron list ``` Review each job and ask: "Is this still providing value?" Pause or remove jobs that are not. There is no `cron pause`: pausing a job is `disable`, and the docs say to "Use `openclaw automations enable ` to resume scheduled runs." ```bash openclaw cron disable ``` --- ## Strategy 6: Monitor Usage You cannot optimize what you do not measure. Regularly check your usage: ### Daily Usage Report There is no top-level `openclaw usage` command and no `--last` flag. `openclaw gateway usage-cost` will "Fetch usage-cost summaries from session logs", and the window is `--days`, which defaults to 30: ```bash openclaw gateway usage-cost --days 1 ``` ``` API Usage (Last 24 Hours) ========================= Total tokens: 45,832 Estimated cost: $1.23 Breakdown: Heartbeat: 12,400 tokens ($0.15) Conversations: 28,200 tokens ($0.89) Cron jobs: 5,232 tokens ($0.19) By model: claude-haiku-4: 32,100 tokens ($0.32) claude-sonnet-4.6: 13,732 tokens ($0.97) ``` ### Weekly Totals There is no documented `--trend` flag. Widen the window instead, and add `--all-agents` to "Aggregate across all configured agents" or `--agent ` for one of them: ```bash openclaw gateway usage-cost --days 7 --all-agents ``` ``` Daily Cost Trend (Last 7 Days): Mon: $1.45 ████████████████ Tue: $0.98 ███████████ Wed: $2.31 █████████████████████████ Thu: $1.12 ████████████ Fri: $0.87 ██████████ Sat: $0.43 █████ Sun: $0.52 ██████ Average: $1.10/day Projected monthly: $33.00 ``` ### Set Up Usage Alerts There is no `openclaw config edit`. Ask the CLI where the live file is: `openclaw config file` "Prints the active config file path", and you open that path in your editor. ```bash openclaw config file ``` ```json { "cost_management": { "alerts": [ { "type": "daily_spend", "threshold": 3.00, "channel": "telegram", "message": "Daily spending exceeded $3. Current: {amount}" }, { "type": "unusual_spike", "threshold_multiplier": 3.0, "channel": "telegram", "message": "Unusual usage spike detected: {details}" } ] } } ``` --- ## Strategy 7: Use Free and Cheaper Providers ### Google Gemini Free Tier Gemini 3 Flash has a generous free tier that covers many heartbeat and lightweight tasks: - 15 requests per minute - 1 million tokens per minute - 1,500 requests per day Configure Gemini as your heartbeat model: ```json { "heartbeat": { "model": "gemini-3-flash" }, "providers": { "google": { "api_key": "your-key", "enabled": true } } } ``` ### Mix Providers for Cost Optimization ```json { "model_routing": { "heartbeat": "gemini-3-flash", "quick_response": "gpt-5.2", "default": "claude-sonnet-4-6-20260217", "complex_reasoning": "claude-sonnet-4-6-20260217" } } ``` This uses the cheapest provider for each task category, potentially reducing costs by 50-70%. --- ## Monthly Cost Estimates by Use Case ### Light Personal Use ($5-15/month) - Haiku default model - 30-minute heartbeat with active hours - 2-3 cron jobs - 10-20 messages per day ### Moderate Personal Use ($15-40/month) - Sonnet 4.6 default, Haiku 4 heartbeat - 15-minute heartbeat with active hours - 5-10 cron jobs - 30-50 messages per day ### Heavy / Professional Use ($40-100/month) - Sonnet 4.6 default, Opus 4.6 for complex tasks - 10-minute heartbeat - 10+ cron jobs - 100+ messages per day ### Unoptimized ($200-500/month) - Opus 4.6 for everything - 10-minute heartbeat, 24/7 - No active hours - No model routing > **Warning** > > The difference between an optimized and unoptimized configuration can be 10-20x in monthly costs. Spend 20 minutes configuring model routing and heartbeat settings, it pays for itself immediately. --- ## Cost Optimization Checklist - [ ] Heartbeat uses Haiku 4 or Gemini 3 Flash (not Sonnet or Opus) - [ ] Heartbeat interval is 15 minutes or more - [ ] Active hours are configured to reduce overnight cycles - [ ] Model routing is configured for different task types - [ ] Smart escalation is enabled - [ ] Context window size is limited - [ ] Spending limits are set on provider dashboards - [ ] OpenClaw cost alerts are configured - [ ] Cron jobs use appropriate models (Haiku 4 for simple tasks) - [ ] Unnecessary cron jobs are paused or removed - [ ] Usage is monitored weekly - [ ] Consolidated cron jobs where possible --- ## Next Steps - **[Heartbeat & Cron Configuration](/docs/heartbeat-config)**: detailed heartbeat tuning - **[API Key Setup](/docs/api-key-setup)**: configure multiple providers for cost optimization - **[Understanding Skills](/docs/skills-explained)**: skills affect cost through API usage - **[Troubleshooting](/docs/troubleshooting)**: when cost limits cause unexpected behavior ## Sources - [OpenClaw: Config](https://docs.openclaw.ai/cli/config) - [OpenClaw: Automations (cron)](https://docs.openclaw.ai/cli/cron) - [OpenClaw: Query a running Gateway](https://docs.openclaw.ai/cli/gateway/query)