Keeping API Costs Under Control with Multiple OpenClaw Agents

The Question Everyone Asks

"Okay, multiple AI agents sound cool. But what does it actually cost?"

The honest answer: less than you think — if you build it right. More than necessary — if you don't.

Here are our real numbers and the decisions behind them.

---

Our Starting Point: 6 Agents, One Bill

We run 6 OpenClaw agents:

Each runs 24/7 on a Hetzner server. Our monthly Anthropic bill: under €500.

That sounds like a lot. Until you calculate what 40+ hours of manual work per week would otherwise cost.

---

Cost Driver #1: The Model

The biggest lever is model selection. The difference between Claude Opus and Claude Haiku isn't a small detail — it's a factor of 15 in price.

| Model | Input (1M tokens) | Output (1M tokens) |

|-----------------------------|-------------------|--------------------|

| claude-opus-4-5 | $15 | $75 |

| claude-sonnet-4-5 | $3 | $15 |

| claude-haiku-3-5 | $1 | $5 |

Choosing the right model for each agent is the most important cost control there is.

Our assignments:

Just through this differentiation, we save ~60% compared to "everyone on Opus."

---

Cost Driver #2: Heartbeat Frequency

OpenClaw can "wake" agents at regular intervals — the heartbeat system. By default, every 30 minutes.

Every heartbeat costs tokens. And when an agent performs unnecessary checks during a heartbeat — querying the email inbox, loading the calendar, starting a web search — it adds up.

What we optimized:

```

// HEARTBEAT.md for Alex (example)

Only act when:

```

Alex now responds to ~80% of all heartbeats with `HEARTBEAT_OK` — no tool call, minimal token consumption.

Rule of thumb: The clearer the conditions in HEARTBEAT.md, the fewer unnecessary actions.

---

Cost Driver #3: Cron Jobs vs. Heartbeats

Not every recurring task should run as a heartbeat. Cron jobs are often more efficient because:

1. They start isolated — no session history overhead

2. They can use an optimizable model (e.g., cheaper model for routine tasks)

3. They only run when there's actually something to do

Bad example (expensive):

Good example (cheap):

That's a 16x difference — for the same task.

---

Cost Driver #4: Context Windows and Session History

OpenClaw loads the session history on every turn. The longer a session, the more input tokens are sent with every turn.

Problem: Long-lived sessions get more expensive over time.

Our solution:

In OpenClaw, there's the `sessionTarget` field in cron jobs for this:

```json

{

"sessionTarget": "isolated",

"payload": {

"kind": "agentTurn",

"message": "Analyze the last 5 GitHub issues and summarize them."

}

}

```

`isolated` = no history overhead, cleaner result, cheaper execution.

---

Cost Driver #5: Tool Calls

Every tool call generates output tokens (for the function call itself) and input tokens (for the result that comes back). Web searches, email reads, file operations — everything has a price.

What helps:

---

Monitoring: What We Actually Track

The OpenClaw UI shows token consumption per session. Additionally we use:

1. Anthropic Console → daily token totals per API key

2. Weekly review → which agent got more expensive? Why?

3. Budget alerts set to daily limits in the Anthropic Console

When an agent suddenly consumes 3× more tokens than the previous week, that's a signal — often an infinite cron job loop, a forgotten tool-call chain, or an unexpectedly long session.

---

Our Cost Summary (Real Numbers)

| Agent | Model | Avg tokens/day | Avg cost/month |

|--------|---------|----------------|----------------|

| Sam | Sonnet | ~80,000 | ~€36 |

| Peter | Sonnet | ~60,000 | ~€27 |

| Maya | Sonnet | ~70,000 | ~€32 |

| Iris | Sonnet | ~90,000 | ~€40 |

| Alex | Haiku | ~50,000 | ~€5 |

| Atlas | Opus | ~30,000 | ~€68 |

Total: ~€208/month — plus server (Hetzner CX21: ~€6/month).

That's well under the €500 budget we set for ourselves. The buffer exists for weeks with more activity or when agents get larger tasks.

---

The Key Takeaways

1. Model selection is the biggest lever — Haiku where it's sufficient, Opus only where it counts

2. Define heartbeat conditions sharply — HEARTBEAT_OK costs almost nothing

3. Cron jobs for recurring tasks — isolated sessions are cheaper

4. Keep an eye on session history — long sessions get more expensive

5. Set up monitoring — budget alerts in the provider console

The detailed configuration — HEARTBEAT.md templates, cron job patterns, Docker Compose with model variables — is documented in the OpenClaw Setup Playbook. Including the monitoring setup we use for our weekly review.

Fully available in German too. 🇩🇪