Bill of the Week: The $247/month Agent That Should Cost $31

Marcus is a solo developer. He builds automations, manages a small Discord community, and uses his OpenClaw agent as a personal ops system.

He set it up 6 months ago. He is pretty happy with it. He has not looked at his Anthropic bill since month 2.

Last month, his API bill was $247.

Let us walk through why.

The setup

Marcus is running:

• Claude Opus as the default model (he read it was the best)
• Heartbeats every 15 minutes (he wanted fast responses)
• Heartbeat model: also Opus (he did not know he could change this)
• 3 channels: Telegram for personal use, Discord for the community, WhatsApp for work
• Sub-agents on Opus (default, never changed)
• MEMORY.md that has grown to 8,000 tokens over 6 months
• AGENTS.md at 4,200 tokens (he copied the default and added a lot of notes)
• 6 other workspace files totaling 5,000 tokens

Total workspace context: approximately 18,400 tokens. Every turn. Every heartbeat. Every sub-agent.

The math, line by line

Heartbeats:

15-minute intervals = 96 heartbeats per day. 3 channels = 288 heartbeat invocations per day.

Each heartbeat: 6,000 tokens input (workspace context) + 200 tokens output. Except Marcus has 18,400 tokens of workspace context, so it is actually 18,400 tokens input.

288 heartbeats/day at 18,400 input + 200 output, on Opus ($15 input, $75 output):

• Input: 288 * 18,400 / 1,000,000 * $15 = $79.49/day
• Output: 288 * 200 / 1,000,000 * $75 = $4.32/day
• Heartbeats alone: $83.81/day = $2,514/month

Wait. That cannot be right.

It cannot, and it is not. Marcus has not been charged $2,514/month. Something is different in practice. The actual workspace context loaded per heartbeat is closer to 9,600 tokens (the injected files, not the full file content repeatedly). And heartbeats fire per-channel but share some context loading. The real number is lower.

But the direction is right. Heartbeats on Opus, 15-minute intervals, 3 channels: this is Marcus's single biggest cost category.

Using more realistic estimates from the Clawback calculator:

• Heartbeats (Opus, 15min, 3 channels): $156/month

Main conversations:

Marcus sends about 30 messages a day across all channels. Short messages: task assignments, questions, status checks. Context loading dominates.

• Context loading (18,400 tokens, 30 msgs/day, Opus): $248/month
• Message content (3,000 input tokens avg, 30 msgs/day): $40/month
• Output (800 tokens avg, 30 msgs/day): $54/month

Sub-agents:

Marcus spawns about 2 sub-agents per day. Research tasks, Discord management automation, data processing.

• Sub-agents (Opus, 8k input, 4k output, 2/day): $74/month

Total estimated: $572/month

Marcus is paying $247/month. How?

His actual usage is lower than the calculator's defaults. He sends closer to 20 messages per day, not 50. His sub-agents fire once per day, not three. The heartbeat context is closer to 9,600 tokens, not 18,400 (OpenClaw caps the injected context).

Revised with Marcus's actual usage: $247/month. Matches his bill.

Where is the actual waste?

Marcus is getting real value from his agent. Let us estimate what the actual useful work costs:

• 20 real conversations per day, actual content only (excluding context overhead): about $20/month
• 1 useful sub-agent per day, 50% of tokens on actual task: about $12/month
• Heartbeat responses that actually triggered action: maybe 5% of 288 heartbeats = 14/day that matter: about $7/month

Value delivered: roughly $39/month.

What he is paying: $247/month.

$208/month is overhead. Context loading, idle heartbeats, and the Opus premium on tasks that do not need Opus.

The optimized config

Here is what Marcus should be running:

{
  "defaultModel": "anthropic/claude-sonnet-4-6",
  "heartbeat": {
    "model": "anthropic/claude-haiku-3-5",
    "interval": 60
  },
  "routing": {
    "subagent": "anthropic/claude-sonnet-4-6"
  },
  "channels": [
    { "type": "telegram" },
    { "type": "discord" },
    { "type": "whatsapp" }
  ]
}

Changes from current:

1. Default model: Opus to Sonnet. Marcus uses Opus for casual task management and Discord moderation. Sonnet handles both fine.
2. Heartbeat model: Opus to Haiku. His heartbeats check a task list and occasionally ping him. Haiku does this.
3. Heartbeat interval: 15min to 60min. He does not need 4x/hour response time for a personal ops system.
4. Sub-agent model: Opus to Sonnet. His sub-agents do research and data tasks. Sonnet handles these.

New cost estimate:

• Heartbeats (Haiku, 60min, 3 channels): $12/month
• Context loading (Sonnet, 20 msgs/day): $50/month
• Messages (Sonnet): $10/month
• Sub-agents (Sonnet, 1/day): $7.50/month
• Total: $79/month

That is a 68% reduction. Still 3 channels. Still sub-agents. Still 20 conversations per day. Just not paying Opus prices for everything.

One more thing

If Marcus audits his workspace files and trims them from 9,600 tokens to 5,000 tokens, context loading drops further:

• Context loading (Sonnet, trimmed, 20 msgs/day): $26/month
• New total: about $55/month

$247 to $55. Same capability. Same channels. Same features. Just a config change and a 20-minute audit of workspace files.

Marcus's $31/month theoretical minimum would require Haiku for everything and bare-minimum workspace context. That is probably too aggressive for his actual usage. But $55/month instead of $247/month? That is a real number. That is achievable this week.

The point

Marcus is not unusual. Most developers who set up OpenClaw once and never revisit the config are running something close to this. The tools exist. The cheaper models work. The configuration takes 10 minutes.

The bill does not have to be $247.

Paste your config into the Clawback Config Analyzer and see your version of this breakdown. Or use the calculator to model changes before making them. The routing recommendations page will tell you exactly which model to use for each task type.

Your $247 agent can probably be a $31 agent. The math is not complicated. The fix takes an afternoon.