OpenClaw pricing: what it actually costs

The license is free. The bill comes from three places

OpenClaw is open source, so the software itself costs nothing whether you run one agent or a thousand. Every real dollar you spend on it lands in one of three buckets.

First, the model: OpenClaw connects to an LLM provider, and every message, heartbeat and context read is tokens. Second, the hosting: the agent is a long-running process that needs a machine that stays on. Third, the operations: setup, security, updates, monitoring and recovery, paid in time or in a managed fee. Most pricing articles only talk about the first bucket, which is exactly why so many people get the total wrong.

Why OpenClaw bills explode

The horror stories are real: Federico Viticci burned through $3,600 in a month, a Reddit user spent $200 in a single day, a developer shared a $623 monthly invoice. But the cause is not the software. It is the billing model.

The default setup uses pay-as-you-go API keys. Anthropic charges $3 per million input tokens and $15 per million output tokens for Claude Sonnet, which sounds cheap until an always-on agent runs crons, re-reads context and processes recurring tasks around the clock. A power user can burn through 100M+ tokens a month, and that is how a setup quietly climbs to $400-800/month. Metered API pricing at scale is brutal, and most users do not need it.

The fix most articles skip: subscription auth

OpenClaw can authenticate with an AI subscription instead of a metered key, and this single change is the difference between a $700 invoice and a $20 one.

• Anthropic: use your Claude Pro or Max subscription token instead of an API key, for flat-rate usage.
• OpenAI: connect your ChatGPT Plus account via Codex OAuth, so usage rides the $20/month you already pay.
• Google: authenticate via Gemini CLI OAuth, covered by a Google One AI Premium subscription.
• GitHub Copilot: an existing Copilot subscription also works with OpenClaw at no extra model cost.
• The math for a heavy user: roughly $400-800/month on metered keys versus $20-100/month on subscriptions, a 5-10x reduction.

When pay-as-you-go keys still win

Subscriptions are not always the answer, and pretending otherwise would be dishonest.

• Light usage: under roughly 10M tokens a month, pay-as-you-go can come out cheaper than a subscription you barely use.
• Specific models: some models are API-only, for example heavyweight reasoning tiers, so subscription auth cannot reach them.
• Teams and volume: business accounts with volume discounts can beat consumer subscription pricing, and give you proper usage attribution per agent, which matters as soon as you bill clients.

Hosting: from a $5 VPS to managed

The second bucket is the machine. Self-hosting on a small VPS costs $5-20/month, and hardware you already own, a Mac mini or a home server, brings the marginal cost close to electricity. That is the visible part.

The four common setups, priced end to end:

• Metered API keys, self-hosted: $50-800+/month depending on usage, plus your ops time. The setup behind every horror story.
• Subscription auth, running on hardware you own: about $20/month total, plus your ops time.
• Subscription auth on a rented VPS: roughly $25-40/month, plus your ops time.
• Managed hosting: a predictable fee that covers the server, monitoring, recovery and updates, with your own subscription or keys on top.

The line item nobody budgets: operations

Self-hosting looks cheapest on paper: a $5 server plus a $20 subscription, call it $25/month. That number ignores the real line item, your time. The setup weekend. The security configuration you have to get right. The crash at 3am. The update that breaks an integration. The monitoring you swear you will add and never do.

At any hourly rate above zero, the cheap option is often the expensive one. For a hobbyist who enjoys infrastructure, the time is the fun part and this argument does not apply. For a business, ops time is payroll, and it scales linearly with the number of agents unless something automates it.

Is there a free way to run OpenClaw?

Technically yes. The software is free; only the model and the machine cost anything, and both have zero-dollar paths.

• OpenRouter's free tier offers limited daily usage at no cost, enough to evaluate seriously.
• Local models via Ollama, such as Llama or Mistral, remove API costs entirely if your hardware can carry them.
• Provider free trials cover initial experimentation.
• Effectively free: if you already pay for Claude Pro, ChatGPT Plus or Copilot, OpenClaw piggybacks on that subscription at no extra model cost. This is the real free option most people miss.

What changes at fleet scale

Everything above is the single-agent picture. Run agents for clients and the question stops being "what is my bill" and becomes unit economics: what does one agent cost per day, how many agents fit safely on one machine, and who carries the on-call.

This is where density decides margins. Most agents idle far below their peak, so a platform that over-provisions safely packs roughly 3x more agents per machine than naive one-box-per-agent hosting, and cost per agent drops accordingly. Molted prices this the way fleets consume it, per instance per day, pro-rated, with 4-tier auto-recovery and per-instance observability included, so the ops bucket stops scaling with headcount. The platform has run in production since January 2026, and the same team operates molted.cloud with 300+ managed clients.

OpenClaw, MoltBot, Clawdbot: same project, same pricing

A quick clarification if the names confuse you: OpenClaw, MoltBot and Clawdbot are the same project, renamed twice for trademark reasons. Old articles about "MoltBot pricing" or "Clawdbot cost" describe the same software and the same cost structure: free license, then model, hosting and ops. Everything on this page applies regardless of which name you found first.