OpenClaw Token Usage: Why It Costs More Than You Expect

New OpenClaw users frequently discover their API bill is significantly higher than anticipated. "I only had 10 conversations today — why did I use 200,000 tokens?" This post explains exactly how OpenClaw uses tokens, why agent conversations are more expensive than direct LLM chats, and what you can do about it.

The Token Multiplication Effect

When you send a message to OpenClaw, the following all count as input tokens:

1. Your message: 20–200 tokens
2. Active skills context: 500–3,000 tokens per installed skill (tool descriptions and examples)
3. Conversation history: Grows with every exchange — 50 messages ≈ 5,000–20,000 tokens
4. Agent memory: Relevant memories retrieved and injected — 500–2,000 tokens
5. System prompt: Your custom instructions — 100–1,000 tokens
6. LLM's response (output): 200–800 tokens (output tokens cost 4–5× more per token)

A "simple" 10-message conversation with 3 active skills can consume 50,000–100,000 tokens in context alone.

Why Skills Are Expensive

When you install a skill, OpenClaw includes its full SKILL.md in every context. A complex skill like Google Workspace includes:

• 12+ tool descriptions with parameters
• Usage examples
• Context about how to chain tools

This adds approximately 2,000 tokens to every single message. With 5 active skills, you're adding 10,000 tokens to each call before your actual message.

Context Window Growth

In a multi-turn conversation, the full conversation history is included in each call. After 20 exchanges:

Message	Tokens Added	Cumulative Context
1	500	500
5	2,000	10,000
10	4,000	30,000
20	8,000	80,000

At Claude 3.5 Sonnet pricing ($3/M input tokens), 80,000 tokens = $0.24 per message near the end of a long conversation.

Automation Costs: The Multiplier Effect

If you set up automations running every hour (common for monitoring tasks), each run includes:

• Full skills context
• Last few hours of memory
• Previous automation outputs stored in context

A "lightweight" hourly automation running 24 times a day can cost $0.50–$2.00/day in API costs alone — $15–$60/month for that single automation.

Real Estimate for Different User Types

User Type	Daily Conversations	Monthly API Cost (Claude)
Light user	5 simple exchanges	$3–$8
Regular user	20 varied conversations	$15–$35
Power user	50+ conversations + automations	$40–$100
Business user	Multiple agents + frequent automations	$100–$500

Reducing Token Usage

1. Disable unused skills — each active skill costs ~2,000 tokens/message
2. Set conversation history limits — cap at 10–20 messages before summarisation
3. Use cheaper models for automations — DeepSeek V3 at $0.27/M input is 11× cheaper than Claude Sonnet
4. Consolidate automations — one daily briefing vs twelve hourly checks
5. Compress memory — regularly ask OpenClaw to summarise and compress its memory

Frequently Asked Questions

Why can't OpenClaw just not include the full history?

The conversation history is what makes OpenClaw contextually aware — it's why it can remember what you discussed three messages ago. Without it, each message would be stateless. The tradeoff is tokens vs context.

Is there a way to see my token usage per conversation?

Yes. Set "debug": {"log_token_usage": true} in openclaw.json to log token counts per API call to the debug log.