What OpenClaw actually costs — and how to cut the bill.

The bill is almost entirely LLM API tokens. The runtime, the gateway, the storage — all free. So the question is just: which API calls are you making, with which model, how often, and how much context per call?

Most people assume their bill is driven by chat volume. It almost never is. The single biggest line item on a default setup is heartbeats — periodic check-ins the gateway sends so the agent can react to scheduled events. They run whether you're using the agent or not.

Drag the sliders to match your situation. The model dropdown uses published 2026 list prices.

The default heartbeat interval is 3 minutes. That's 480 calls per day, ~14,400 per month. Each one consumes 1–2k input tokens (the agent's running context). On Sonnet 4.6, that's ~$15/mo on heartbeats alone — before any actual work.

The fix is two-part:

• Route heartbeats to a cheap model. Haiku, GPT-4o-mini, or a local Ollama model. Quality doesn't matter; this is just a "is anything pending?" check.
• Increase the interval. If your agent doesn't have scheduled jobs, 15 minutes is fine. If it does, match the heartbeat to the cron resolution.

{
  "heartbeat": {
    "intervalMinutes": 15,
    "model": "claude-haiku-4-5"
  }
}

The single most impactful change after fixing heartbeats: stop using the same flagship model for everything. 70–80% of an agent's calls don't need flagship-quality reasoning.

Full routing setup including local models in the Ollama guide.

Anthropic's prompt cache gives you 0.1x input pricing on repeated content. OpenClaw loads the same MEMORY.md and system prompt every turn — those are perfect cache candidates.

{
  "anthropic": {
    "promptCaching": true,
    "cacheBreakpoints": ["system", "memory"]
  }
}

In benchmarks, this drops monthly bills 30–50% by itself for memory-heavy agents. The cache TTL is 5 minutes, so it works best when the agent is actively used (multiple turns close together) rather than once-an-hour pings.

Every file you load into the agent's context costs money. Most people over-load context because it's easy and the cost isn't visible per-call.

• Don't dump everything into MEMORY.md. Use memory_search for the things the agent needs occasionally; reserve MEMORY.md for facts loaded every session.
• Trim system prompts. If your prompt is 4k tokens, half of that is probably redundant.
• Compact aggressively. The default compaction threshold is generous; bring it down to 50% for active sessions to keep average context smaller.
• Use vector recall over file dumps. Loading a 50k-token doc into context vs searching it for the 500 relevant tokens is a 100x cost difference.

Anthropic's Batch API gives you 50% off both input and output if you can wait up to 24 hours for the result. OpenClaw cron jobs that don't need immediate output should use it.

{
  "schedule": "0 0 * * *",
  "task": "Compile daily digest",
  "options": { "batch": true, "deliverBy": "09:00" }
}

Two quotas, two layers. The first is OpenClaw's soft cap — it warns and pauses when hit, but a misbehaving agent can route around it. The second is the provider's hard cap on the API key itself, which stops calls dead.

{
  "quotas": {
    "maxMonthlySpend": 50,
    "warnAt": 0.8,
    "pauseAt": 1.0
  }
}

Then in the Anthropic console: Settings → API keys → spending limit. Set it 50% above your soft cap as a safety net.