Understanding OpenClaw Token Consumption | Generated by AI

Home 2026.05 English 中文

---

Question: Why does OpenClaw consume ~3 million tokens per day even when I’m not actively using it (as seen in OpenRouter)?

Answer:

Short answer: it’s almost certainly still doing work in the background.

When a system like OpenClaw (or similar agent-style frameworks) is running, “doing nothing” from your perspective doesn’t mean idle. These systems are often built as continuous loops or autonomous agents, and they can burn tokens quietly through several mechanisms:

---

1. Continuous agent loop (most common cause)

Many agent frameworks run something like:

while True:
    observe()
    think()
    act()

Even if no meaningful task is happening, the model is still:

• re-evaluating state
• generating thoughts/plans
• calling the LLM repeatedly

Each loop = tokens consumed.

If the loop runs fast (e.g., every 1–5 seconds), you can easily hit millions of tokens/day.

---

2. Hidden background tasks

OpenClaw-style systems often include:

• memory summarization
• log analysis
• environment polling (GUI, browser, etc.)
• retry loops for failed tasks

These can trigger LLM calls like:

• “Summarize current state”
• “Decide next action”
• “Check if task complete”

Even if nothing changes, the model is still queried.

---

3. Long context accumulation

If the system keeps appending history:

• each new call includes previous messages
• token usage grows over time

Example:

• 1k tokens → 2k → 4k → 8k per call Now even a few calls = huge daily usage

---

4. Multiple parallel agents or threads

Some frameworks silently run:

• planner agent
• executor agent
• evaluator agent

So one “step” might actually be 3–5 LLM calls.

---

5. Retry / failure loops

If something is slightly broken:

• API call fails
• action doesn’t complete

The system may retry endlessly:

try → fail → reflect → retry → fail → reflect ...

This can explode token usage very quickly.

---

6. OpenRouter accounting makes it visible

Platforms like OpenRouter show total token usage, so you suddenly notice:

• background loops
• inefficiencies
• runaway agents

Without that visibility, you’d never realize it’s happening.

---

Rough intuition

3 million tokens/day ≈

• ~35 tokens/sec continuously, OR
• ~1,500 tokens per call × 2,000 calls/day

That’s totally consistent with a fast loop + moderate context.

---

What to check immediately

1. Is the process still running?

   • Background script / server / Docker container?

2. Loop frequency

   • Look for while True or polling intervals

3. Context size

   • Are you appending full history every call?

4. Number of LLM calls per step

   • Planner + executor + critic?

5. Retry logic

   • Any infinite retry loops?

6. Logs

   • You’ll likely see constant LLM requests even when idle

---

Practical fixes

• Add a sleep interval (e.g., 5–30 seconds)
• Stop loop when no task exists
• Limit context length (truncate history)
• Disable unused agents
• Add a hard token budget per day
• Log every LLM call (you’ll instantly see the culprit)

---

Key insight

Agent systems are not like normal apps. If you don’t explicitly stop them, they behave like:

“Always thinking, even when there’s nothing to think about.”

And thinking = tokens = money.

---

If you want, you can paste your OpenClaw config or loop code, and I can pinpoint exactly where the tokens are being burned.

---

Back Donate