OpenClaw Token Usage & Cost Control

OpenClaw agents can become expensive when they use long prompts, large tool outputs, growing memory files, browser actions, and repeated model calls. Cost is not just about which model you picked. It depends on how the agent uses tokens during real workflows.

This guide shows you how to see where your tokens go, set hard limits before things spiral, and either control pay-per-token costs carefully or switch to a flat-rate setup that takes the guesswork out entirely.

A token is the unit AI providers use to measure API usage. Plain English:

• One token is roughly 4 English characters or about 3/4 of a word
• A 1,000-word document is usually around 1,300 tokens
• Providers usually charge per 1 million tokens, also called MTok

A quick example:

This distinction matters a lot for your bill.

• Input tokens - prompt, system message, conversation history, files, tool results, memory
• Output tokens - model reply, reasoning steps, generated plans, summaries, code, final answers

In many paid APIs, output tokens cost more than input tokens. Long generated replies often increase cost faster than long prompts. Capping how much your agent writes saves money quickly.

The main drivers behind unexpected bills:

• Long conversation history sent with every message
• Large browser or tool outputs added to context
• Repeated failed retries silently doubling work
• Expensive default model used for everything
• Too many always-on workflows running in the background
• Large files or web pages added to context unnecessarily
• No spending limits set at the provider
• No model routing - every task hits your premium model

Most OpenClaw cost problems do not come from one big request. They come from small inefficient workflows running again and again.

Math you can do in your head:

More practical version with real numbers:

Example with a $3 input / $15 output model:

Commands can vary by OpenClaw version. Check the built-in help:

• OpenAI usage dashboard
• Anthropic console
• Google Cloud / AI Studio usage
• OpenRouter dashboard
• Groq console

Use separate API keys for different workflows. Provider dashboards then break down usage by key:

Want predictable costs without tracking dashboards?

Ampere.sh Pro is flat-rate managed OpenClaw hosting. Pooled API access, smart routing, one bill, no spreadsheets.

Do not wait until the end of the month to check what your agent cost you. Set hard limits at the provider level so spending physically cannot exceed your budget.

• Set provider-level spending limits, not just internal notes
• Add soft alerts at 50% and 80% of your budget
• Use prepaid credits where possible - they cannot overflow
• Disable unused API keys to prevent forgotten workflows from running

Spend two minutes each morning checking if today's usage looks normal. Big jumps usually mean something broke in a workflow.

Check 7-day trends in your provider dashboards. Patterns hide in 24-hour windows but jump out across a week.

Use separate keys per major workflow. When usage spikes, the right key tells you exactly which workflow to investigate.

Retries multiply cost silently. Watch logs for repeated failures:

If output tokens are rising faster than input tokens, your agent's responses are getting longer. Cap them in your prompts.

Each is small. Stacked together, they cut most bills by half:

• Shorten system prompts
• Keep SOUL.md, AGENTS.md, and memory files clean
• Summarize old conversations instead of dragging full history
• Limit browser output to the relevant section
• Avoid pasting full documents when only one section is needed
• Use cheaper models for simple tasks
• Use stronger models only for complex work
• Cap response length in your instructions
• Reduce failed retries with better error handling
• Remove unused workflows
• Use prompt caching when supported (Anthropic, OpenAI)

Quick example of switching to a cheaper default model:

Use the exact model name supported by your provider. See the cost reduction guide for more.

Use Ampere.sh if you want:

• Predictable monthly cost
• Managed OpenClaw hosting without server work
• Less API cost tracking
• Less server maintenance
• Always-on workflows that just run
• Easier provider setup
• Fewer gateway, port, and uptime issues

Try Ampere.sh if you want OpenClaw running with predictable hosting instead of watching token dashboards every week.

Watch for these patterns. They mean you're either losing money already or about to:

• Daily usage doubled suddenly
• One workflow uses most of the tokens
• Output tokens are consistently high
• Conversation history is too large
• Browser automation pulls huge pages
• Logs show repeated failed retries
• Rate limits appear often
• Monthly bill is rising without new workflows
• Expensive model is set as default for all tasks

If you take only a handful of things from this guide, take these:

• Track token usage early using the configuration guide
• Set hard provider budgets, not just notes
• Use cheaper models by default
• Route expensive models only to complex work
• Clean memory files and system prompts regularly
• Watch tool outputs and retry loops
• Use Ampere.sh if predictable managed hosting matters more than manual token tracking