Cost Management — Running OpenClaw Without Going Broke

OpenClaw is free and open source. The software costs nothing. But the AI models it calls? Those cost money. And if you are not thoughtful about how you configure things, you can burn through $50-100+ in a week without realizing it.

This is the number one surprise for new OpenClaw users. They set up their assistant, start using it enthusiastically, and then get an API bill that makes their stomach drop. It does not have to be that way. With the right configuration, most people can run a highly capable OpenClaw assistant for $15-40 per month in total API costs.

This page covers everything you need to know about where the money goes and how to control it.

Where the Money Goes

Before you can optimize costs, you need to understand the cost structure. OpenClaw does not charge you anything — you pay the AI providers directly for every API call your assistant makes.

Here is a breakdown of the cost categories:

The dominant cost is LLM calls. That is where your optimization efforts should focus.

Model Selection Strategy

The single most impactful thing you can do to control costs is use the right model for the right task. Not every message needs the most powerful (and expensive) model.

Model Pricing Overview

Prices change frequently, but here are approximate costs per 1 million tokens as of early 2025 (check current pricing at each provider’s site):

The price difference between the top and bottom of this list is 100x or more. Using Claude Opus for a task that GPT-4o-mini could handle is like hiring a brain surgeon to put a bandage on a paper cut.

The Tiered Model Approach

Configure your OpenClaw to use different models for different task complexities:

models:
  primary: "claude-sonnet-4-20250514"          # Default for most tasks
  reasoning: "claude-opus-4-0-20250116"          # Complex analysis, important decisions
  simple: "gpt-4o-mini"                  # Quick questions, simple tasks
  sub_agents: "gpt-4o-mini"             # Background worker agents
  summarization: "gemini-1.5-flash"      # Summarizing long documents

Then in your SOUL.md, give your agent guidance on when to escalate:

## Model Selection
 
Use your default model (Sonnet) for most tasks. Escalate to the reasoning model
(Opus) only when:
- The task requires multi-step logical reasoning
- I explicitly ask for deep analysis
- The task involves important decisions (financial, career, health)
- Creative writing that needs to be exceptional quality
 
Use the simple model (GPT-4o-mini) when:
- Answering factual questions with clear answers
- Simple formatting or restructuring text
- Classification tasks ("Is this email urgent?")
- Summarizing short pieces of text

Real-World Impact

Here is what the tiered approach looks like in practice for a typical day with 50 interactions:

Over a month, the tiered approach saves $20-150 compared to using a premium model for everything.

Model Failover Chains

What happens when your primary model’s API is down? Without a failover chain, your assistant just… stops working. With one, it gracefully switches to an alternative.

Setting Up Failover

models:
  failover:
    - provider: "anthropic"
      model: "claude-sonnet-4-20250514"
    - provider: "openai"
      model: "gpt-4o"
    - provider: "google"
      model: "gemini-1.5-pro"
    - provider: "openai"
      model: "gpt-4o-mini"     # Last resort: cheap but always available

The Gateway tries each model in order. If Anthropic’s API is down, it falls through to OpenAI, then Google. The last entry should always be your cheapest, most reliable option — the model that might not be the best, but is almost never down.

Failover and Cost

A subtle benefit of failover chains: they can also be used for cost-based routing. Some providers offer the same model quality at different prices, or different models with similar capabilities at different price points. Your failover chain can be ordered by cost preference, not just availability.

OAuth vs. API Keys — A Hidden Cost Saver

Here is something many new users miss: not everything costs tokens.

When OpenClaw accesses Google Calendar, Gmail, Google Drive, or other services via OAuth, those calls do not go through your LLM API. They are direct service API calls. The LLM model is only involved in deciding what to do and processing the results — not in the actual data retrieval.

The practical takeaway: set up OAuth integrations wherever possible. Reading your calendar through OAuth is nearly free. Having your LLM search the web for your calendar information is expensive and less reliable.

Token Monitoring

You cannot optimize what you do not measure. OpenClaw provides several ways to track your token usage.

Gateway Logs

The Gateway logs every model call with token counts:

# View recent model calls with token usage
tail -f ~/.openclaw/logs/gateway.log | grep "tokens"

A typical log entry looks like:

[2025-01-15 14:23:01] model=claude-sonnet-4-20250514 input_tokens=1,247 output_tokens=892 cost=$0.0156

Dashboard Metrics

If you are using the web client at http://localhost:18789, the dashboard shows:

• Total tokens used today/this week/this month
• Cost breakdown by model
• Cost breakdown by task type (chat, cron, sub-agents)
• Most expensive sessions

Provider Dashboards

Check your spending directly at each provider:

• Anthropic: console.anthropic.com → Usage
• OpenAI: platform.openai.com/usage
• Google: aistudio.google.com → Usage

Setting Spending Alerts

Most providers let you set spending limits or alerts:

Anthropic:

• Set a monthly spend limit in the console
• Receive email alerts at threshold percentages

OpenAI:

• Set hard spending caps
• Configure email notifications

Google:

• Set budget alerts in the Google Cloud console

Set alerts at 50% and 80% of your monthly budget. This gives you time to adjust before hitting your limit.

The Hidden Costs

LLM tokens get all the attention, but several other services quietly add to your bill.

Text-to-Speech (ElevenLabs)

Every time your assistant speaks aloud, you pay ElevenLabs for the audio generation. This cost is character-based, not token-based.

A typical spoken response is 200-500 characters. If your assistant speaks 50 responses per day, that is 10,000-25,000 characters per day, or 300,000-750,000 per month. The Starter plan (30K characters) would not cover daily voice use. The Creator plan (100K) handles moderate use.

Cost reduction tips:

• Use auto_speak: "talk_mode_only" so TTS only activates during voice conversations
• Set max_speak_length: 300 to limit how much text gets converted to speech
• For long responses, have the agent speak a summary and deliver the full text via your channel

Image Generation

If your assistant generates images (DALL-E, Midjourney), each image has a significant cost:

A single image is cheap. But if you have a workflow that generates 10 images for a presentation, that is $0.40-0.80 per run.

Cost reduction tip: Use DALL-E 2 ($0.020 per image) for drafts and previews. Only use DALL-E 3 for final outputs.

Web Search

Many skills that search the web incur costs beyond the LLM tokens. The search API itself may have a cost, and the LLM needs to process the search results (which can be token-heavy).

Cost reduction tip: Be specific in search instructions. “Find the current price of Bitcoin” is one search call. “Research everything about Bitcoin” might trigger dozens of searches, each returning thousands of tokens of results.

Embeddings (Memory System)

OpenClaw’s memory system uses embeddings to find relevant memories. Each memory retrieval involves an embedding API call.

Embedding costs are very low (OpenAI’s text-embedding-3-small costs $0.02 per 1M tokens), but they add up if your memory system is doing hundreds of retrievals per day.

Practical Tips for Reducing Costs

Here are concrete actions you can take today to lower your OpenClaw spending:

1. Audit Your Cron Jobs

Cron jobs run whether you are using your assistant or not. Each cron job triggers a model call. Review your scheduled tasks and ask:

• Does this need to run hourly, or would daily suffice?
• Does this need the primary model, or would a cheap model work?
• Is this cron job actually providing value, or did I set it up to test and forget about it?

# Expensive: Hourly news summary using Sonnet
cron:
  - schedule: "0 * * * *"    # Every hour
    model: "claude-sonnet-4-20250514"
    task: "Summarize top news"
 
# Cheaper: Twice-daily news summary using Flash
cron:
  - schedule: "0 8,18 * * *"  # 8 AM and 6 PM
    model: "gemini-1.5-flash"
    task: "Summarize top news"

The cheaper version runs 12x less often and uses a model that is 20x cheaper. Combined: ~240x cost reduction for that single cron job.

2. Shorten System Prompts

Your SOUL.md, USER.md, and other system context get sent with every single message. If your SOUL.md is 5,000 tokens, that is 5,000 input tokens on every interaction before the user even says anything.

• Keep SOUL.md concise. Aim for under 2,000 tokens.
• Move reference information to files the agent can look up on demand, rather than including it in the system prompt.
• Use conditional context — only load relevant sections based on the task type.

3. Limit Context Window Usage

Long conversations accumulate tokens. A 50-message conversation thread might have 20,000+ tokens of history, all sent with every new message.

sessions:
  max_context_messages: 20      # Only send last 20 messages
  summarize_after: 15           # Summarize older messages instead of including verbatim

This keeps your context window lean. The agent still has access to older context through summaries, but you are not paying full token prices for messages from an hour ago.

4. Use Sub-Agent Model Routing

When spawning sub-agents, always specify the cheapest appropriate model:

## Sub-Agent Cost Rules (in SOUL.md)
 
When spawning sub-agents:
- Research/search tasks → use gpt-4o-mini
- Summarization → use gemini-1.5-flash
- Writing/creative tasks → use your default model (Sonnet)
- Only use Opus when I explicitly request "deep analysis" or "thorough review"

5. Batch Similar Requests

Instead of asking five separate questions (five model calls), batch them into one:

Expensive (5 separate calls):

“What’s the weather?” … “What’s on my calendar?” … “Any urgent emails?” … “Top news?” … “What should I focus on today?”

Cheaper (1 call):

“Give me my morning briefing: weather, calendar, urgent emails, top news, and a suggested focus for today.”

The batched version uses one model call instead of five. The input tokens are similar, but you save on the overhead of five separate API round trips.

6. Disable Unused Features

If you are not using voice, turn off TTS and STT. If you are not using image generation, remove those skills. Every enabled feature is a potential cost center.

7. Set Hard Limits at the Provider Level

As a safety net, set maximum monthly spend limits on each provider account. If something goes wrong — a cron job loops, a sub-agent goes haywire — the spending cap prevents a surprise $500 bill.

Monthly Budget Templates

Here are realistic budget targets for different usage levels:

Light Usage ($10-15/month)

• 10-20 messages per day
• Tiered model selection (mostly GPT-4o-mini and Flash)
• No voice
• 2-3 cron jobs (daily, using cheap models)
• No image generation

Moderate Usage ($20-40/month)

• 30-50 messages per day
• Tiered model selection (Sonnet as primary, mini for sub-agents)
• Voice enabled (ElevenLabs Starter plan)
• 5-10 cron jobs
• Occasional image generation

Heavy Usage ($50-80/month)

• 50-100+ messages per day
• Sonnet as primary, Opus for complex tasks
• Voice enabled (ElevenLabs Creator plan)
• 10+ cron jobs
• Regular image generation
• Phone calls via Twilio

Enterprise/Power Usage ($100+/month)

• Constant usage throughout the day
• Multiple sub-agent workflows
• Heavy voice and phone usage
• Real-time monitoring and alerts
• Multiple users/channels

The Cost Over Time

One last thing to keep in mind: AI model costs have been dropping consistently. Models that cost $15/million tokens in 2024 cost $3 in 2025. The trend shows no sign of slowing.

This means two things:

1. Costs that feel high today will feel reasonable in 6-12 months as providers release cheaper, more capable models
2. Optimizing now teaches you habits that will serve you well even as prices drop — because you will naturally scale up your usage to match the savings

The goal is not to minimize spending at all costs. It is to spend consciously — putting your budget toward the interactions that matter most and using efficient models for everything else.

What’s Next

• Sub-Agents — Learn how to route sub-agents to cheap models
• Deployment — See total cost of ownership including hosting
• Voice & Talk Mode — Understand the voice cost stack
• SOUL.md — Write an efficient system prompt that does not waste tokens