How Much Does OpenClaw Cost? The Real Numbers

The baseline: context loading

Every time you send a message, OpenClaw loads your workspace before the model reads your input. This is not optional. It is how the system works.

What gets loaded on each turn:

• SOUL.md — persona file, roughly 1,200 tokens
• AGENTS.md — workspace rules, roughly 2,000 tokens
• MEMORY.md — long-term memory, grows over time, 1,500+ tokens
• USER.md — info about you, 300 tokens
• HEARTBEAT.md — task checklist, 200 tokens
• TOOLS.md — environment notes, 400 tokens
• Skill descriptions — 8 to 10 installed skills as XML, 1,500 tokens
• System prompt + runtime context — 2,500 tokens

Total: approximately 9,600 tokens. On every single turn. Before you type anything.

Claude Opus 4 input costs $15 per million tokens.

9,600 tokens × $15/MTok = $0.144 per turn in context overhead alone.

At 50 messages per day, that is $7.20/day, or $216/month, just from loading the same files repeatedly.

Heartbeats: the daily background cost

OpenClaw heartbeats are periodic checks. The default is every 30 minutes. That is 48 heartbeats per day.

Each heartbeat loads context (6,000 tokens) and outputs a short response (200 tokens). On Claude Opus 4:

• Input: 6,000 × 48 = 288,000 tokens/day = $4.32
• Output: 200 × 48 = 9,600 tokens/day = $0.72
• Daily heartbeat cost: $5.04
• Monthly: $151 per channel

Two channels (Telegram + Discord)? That is $302/month for your agent to periodically wake up, check its notes, and say HEARTBEAT_OK.

Switch heartbeats to Claude Haiku 4 ($0.80/MTok input) and the same 48 daily heartbeats cost about $0.24/day, or $7.30/month. That is a 95% reduction for identical behavior.

Sub-agents: multiplied context costs

When OpenClaw spawns a sub-agent, that is a new API call with its own context window. Typical sub-agent: 8,000 tokens input, 4,000 tokens output.

At 3 sub-agents per day on Opus:

• Input: 8,000 × 3 = 24,000 tokens = $0.36
• Output: 4,000 × 3 = 12,000 tokens = $0.90
• Daily sub-agent cost: $1.26
• Monthly: $37.80

Sub-agents doing real work (code review, research, writing) produce value. Sub-agents doing simple tasks on Opus pricing do not.

Model pricing table

All prices are per million tokens as of early 2025.

The context/turn column is what you pay just for loading workspace files. Before your actual message costs anything. The difference between Opus and Haiku at this one number is 19x.

Channels multiply everything

OpenClaw supports multiple channels: Telegram, Discord, WhatsApp, and others. Each channel runs an independent session. That means heartbeats per channel, context loads per channel, and costs per channel.

Two channels do not double your cost if you are careful. They absolutely will double your cost if you use the same model and heartbeat frequency on both.

Most users only need one or two channels. Having five connected channels is almost never worth the cost unless each one serves a meaningfully different purpose with meaningfully different workloads.

A real monthly bill at different configurations

What you can do right now

Three changes cover 80% of the savings for most users:

1. Switch heartbeat model to Claude Haiku 4. Heartbeats do not need intelligence. They need to check a list and maybe send a ping.
2. Switch sub-agent model to Claude Sonnet 4. Sonnet handles 95% of real-work sub-agent tasks. The 5% where you need Opus, you will know.
3. Trim your workspace files. Every kilobyte removed cuts costs on every turn, every heartbeat, every sub-agent. It compounds fast.

These are configuration changes. They take 10 minutes. They do not change what your agent can do.

The hidden multiplier: time

OpenClaw is always on. That is the point. But it means the meter is always running. A $5/day agent that you ignore over a weekend is not $5. It is $15 before you notice. An agent running on Opus with 15-minute heartbeats does not cost you when you are sleeping. It costs you whether you are sleeping or not.

Set a budget alert on your Anthropic account. Check your usage dashboard once a week for the first month. The cost surprises almost everyone the first time.