OpenClaw Pricing Guide: Models, Tokens, and Hidden Costs

The pricing model: pay per token

There is no OpenClaw subscription. No per-seat fee. No tier. OpenClaw is open-source software. The costs come entirely from the AI API you configure it to use.

Token pricing works like this: you pay separately for input tokens (what you send to the model) and output tokens (what the model returns). Input is always cheaper. Output costs 3 to 5 times more per token.

A token is roughly 0.75 words. A 1,000-word document is about 1,330 tokens. A full AGENTS.md file is 1,500 to 2,500 tokens depending on how much you have written.

Every API call charges you for the full context window. The model cannot selectively read your files. If AGENTS.md is in the context, you pay for every token in AGENTS.md, every time.

Why pay-per-token is different from SaaS pricing

Most SaaS products have predictable costs. You pay $50/month, you know what you get. The price does not change based on how long your README is.

Pay-per-token pricing does not work that way. Your costs depend on:

• Which model you use
• How many messages you send per day
• How long your workspace files are
• How often heartbeats run
• How many channels you have connected
• How many sub-agents get spawned
• How long the model's responses are

Change any of those factors and your bill changes. Add a detailed MEMORY.md and your bill goes up slightly on every turn. Set heartbeats to 15 minutes instead of 60 and your bill quadruples for that category. It is dynamic and it is invisible until you calculate it.

The hidden costs nobody tells you about

Context loading overhead.

OpenClaw injects workspace files, system prompts, skill descriptions, and runtime context into every message. This is the largest cost for many users. A typical install loads 9,600 tokens per turn. On Opus, that is $0.144 per message before you type anything. At 50 messages/day: $4.32/day, $130/month, just from overhead.

Heartbeat accumulation.

30-minute heartbeats run 48 times per day. Each heartbeat is a full API call: context load, model inference, output. On Opus with 2 channels, heartbeats alone run $300/month. Users almost never factor this in when they set up their agent.

Sub-agent chaining.

Sub-agents are spawned to handle delegated tasks. Each sub-agent is a separate API call with its own context window. A task that spawns 3 sub-agents costs 4x what a single-call task would cost (main call + 3 sub-agents). Complex orchestration workflows multiply costs fast.

Response length.

Output tokens cost 3 to 5x more than input tokens. A verbose model that responds with 2,000 words when 400 would suffice is burning money. You can reduce output length with explicit instructions in your SOUL.md, but most users never do.

The upside of pay-per-token

No monthly minimum. If your agent sits idle for a week, you pay almost nothing. If you need to scale down, just reduce heartbeat frequency or switch to a cheaper model. There are no contracts, no commitments, no cancellation fees.

You also get full control over the cost/quality tradeoff. Use Opus when you need it. Use Haiku when you do not. Most SaaS products do not give you that dial.

For light users (10 to 20 messages/day, hourly heartbeats, Sonnet as the main model), costs are genuinely low. A well-configured minimal setup runs $5 to $15/month.

Budgeting for an always-on agent

The key insight: most of the cost is from the agent running, not from you using it. Heartbeats run around the clock. That makes your bill less sensitive to usage patterns than you might expect.

A useful mental model: think of your OpenClaw cost as two buckets.

• Fixed daily cost. Heartbeats + idle sessions. This runs whether you interact with the agent or not. Calculate this first.
• Variable per-message cost. Context loading + your actual messages + any sub-agents spawned. This scales with usage.

Example on Sonnet 4 with hourly heartbeats on 1 channel:

• Fixed daily: 24 heartbeats × (6,000 + 200 tokens) × $3/$15 per MTok = $0.46/day = $14/month
• Variable: 20 messages × (9,600 + 2,000 context + response tokens) × $3/$15 = roughly $0.80/day = $24/month
• Total: $38/month

The biggest lever: model selection

Model choice dominates everything else. The difference between running Opus everywhere vs. a smart routing strategy is often 10x in cost.

Same agent. Same capabilities. Different cost based entirely on which model handles which task.

Practical budgeting tips

1. Set a cost alert in Anthropic Console. Go to Usage settings and set a soft limit. You will get an email when you hit it. Do this on day one.
2. Check usage after the first week. The pattern of costs is usually surprising. Heartbeats are the biggest shock for most people.
3. Run the heartbeat model on Haiku. This single change saves 90%+ of heartbeat costs. Heartbeats do not need Opus. They need to read a checklist and maybe send a message.
4. Audit your workspace files monthly. MEMORY.md grows over time. Every 500 tokens you trim saves you 1.5 cents per message. At 50 messages/day, that is $22.50/month per 500 tokens saved.
5. Use cron for scheduled tasks, not heartbeats. Cron jobs can run with minimal context. Heartbeats load everything. If you have scheduled work that does not need full workspace context, configure it as a cron job instead.

Is OpenClaw worth the cost?

That depends entirely on what you use it for. An agent that saves you 2 hours of work per week is worth $100/month if your time is worth $50/hour. One that mostly just monitors your inbox and sends you summaries might not justify $150/month.

The question is not whether the cost is high. It is whether the value per dollar is there. Most users who configure their agent well and route models intelligently find the cost completely reasonable. Users who leave everything on defaults and never check their API spend tend to find the cost surprising and not proportional to value.

Configuration is the product. The models are available to everyone. How you wire them together is what determines whether your OpenClaw setup is a bargain or a money pit.