The Honest Answer: It Depends on How You Run It
OpenClaw is open-source software — the code itself is free. But running an AI agent costs money regardless of the platform: you need compute to run it and API credits to power the AI model. The real question is how much, and whether it's worth doing yourself or using a managed service.
This guide breaks down the actual costs for three types of users: light users, power users, and businesses. No hidden fees, no asterisks.
Cost Component #1: The AI Model (API Credits)
This is usually the largest cost, and it's the same whether you self-host or use a managed platform. Every message your AI agent processes costs tokens, and tokens cost money.
Here's what the major models cost through OpenRouter (a unified API gateway) in March 2026:
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Typical Monthly Cost |
|---|---|---|---|
| DeepSeek V3.2 | $0.14 | $0.28 | $1 - $5 |
| Claude 3.5 Haiku | $0.80 | $4.00 | $5 - $20 |
| GPT-4.1 Mini | $0.40 | $1.60 | $3 - $15 |
| Claude Opus 4.6 | $15.00 | $75.00 | $30 - $150 |
| GPT-5.2 | $10.00 | $30.00 | $20 - $100 |
The "Typical Monthly Cost" column assumes 20-50 messages per day with average-length conversations. Light users on budget models might spend $1-5/month. Heavy users on premium models could spend $100+.
Key insight: Your choice of model has a bigger impact on cost than anything else. Using DeepSeek V3.2 instead of Claude Opus 4.6 can reduce API costs by 95% — and for many use cases, the cheaper model is perfectly adequate.
Cost Component #2: Hosting (Where the Agent Runs)
This is where self-hosting and managed hosting diverge significantly.
Option A: Self-Hosting on a VPS
To run OpenClaw yourself, you need a Linux VPS with Docker. Here's what that costs in 2026:
| Provider | Specs | Monthly Cost |
|---|---|---|
| Hetzner CX22 | 2 vCPU / 4GB RAM | $5.29 |
| DigitalOcean Basic | 2 vCPU / 4GB RAM | $24.00 |
| Vultr Cloud Compute | 2 vCPU / 4GB RAM | $24.00 |
| AWS Lightsail | 2 vCPU / 4GB RAM | $36.00 |
| Oracle Cloud Free Tier | Arm 4 vCPU / 24GB RAM | $0 (limited availability) |
You'll need at least 2GB of RAM for the OpenClaw container (it uses Node.js with Chromium for web browsing), plus overhead for the OS and Docker. A 4GB RAM VPS is the sweet spot for running one agent comfortably.
Hidden costs of self-hosting:
- Domain name: $10-15/year if you want a custom gateway URL.
- SSL certificate: Free with Let's Encrypt, but you need to set it up and maintain auto-renewal.
- Backups: $1-5/month for automated VPS snapshots. Optional but strongly recommended.
- Time: Initial setup takes 1-4 hours depending on your Linux experience. Ongoing maintenance (updates, troubleshooting, monitoring) averages 1-2 hours per month.
Option B: OpenClaw Launch (Managed)
OpenClaw Launch handles all the hosting, so you don't need a VPS or any technical setup:
| Plan | Monthly Cost | What You Get |
|---|---|---|
| Lite | $3/mo (first month), then $6/mo | 1 AI agent, Telegram + Discord + WhatsApp, all AI models, all skills, dedicated container |
| Pro | $20/mo | Everything in Lite + multiple agents, priority support, higher resource limits |
What's included: Server hosting, Docker management, automatic updates, HTTPS, container isolation, E2E encryption, web gateway, monitoring, and the configuration dashboard. You supply your own API key (BYOK) for the AI model.
Total Cost by User Type
Let's put it all together for three realistic personas.
Light User: Personal AI Assistant
You want a Telegram bot for personal use — asking questions, getting writing help, occasional web searches. Maybe 10-20 messages per day.
| Component | Self-Hosted | OpenClaw Launch |
|---|---|---|
| AI model (DeepSeek V3.2) | $1 - $3/mo | $1 - $3/mo |
| Hosting | $5 - $24/mo (VPS) | $3 - $6/mo |
| Time investment | 2-4 hrs setup + 1-2 hrs/mo | 10 minutes setup |
| Total | $6 - $27/mo + your time | $4 - $9/mo |
For light users, OpenClaw Launch is actually cheaper than most VPS options and saves hours of technical work. The only scenario where self-hosting wins on cost is Oracle Cloud's free tier (if you can get a spot).
Power User: Daily Driver with Premium Models
You use your AI agent heavily — 50+ messages per day, multiple skills enabled, a premium model like Claude Opus 4.6 for important tasks and a budget model for routine queries.
| Component | Self-Hosted | OpenClaw Launch |
|---|---|---|
| AI models (mixed) | $20 - $60/mo | $20 - $60/mo |
| Hosting | $5 - $24/mo (VPS) | $6 - $20/mo |
| Time investment | 4-8 hrs setup + 2-3 hrs/mo | 10 minutes setup |
| Total | $25 - $84/mo + your time | $26 - $80/mo |
At this level, the hosting cost is a small fraction of total spend — API credits dominate. The main advantage of managed hosting is time savings and reliability. No SSH debugging at midnight when your bot goes down.
Business: Customer-Facing Agent
You run a small business with a WhatsApp AI agent handling 100-200 customer messages per day. You need reliability, fast response times, and multi-language support.
| Component | Self-Hosted | OpenClaw Launch |
|---|---|---|
| AI model (Claude 3.5 Haiku) | $15 - $40/mo | $15 - $40/mo |
| Hosting | $10 - $36/mo (VPS) | $20/mo (Pro plan) |
| Time / ops | 8+ hrs setup + 3-5 hrs/mo | 30 minutes setup |
| Monitoring / uptime | DIY (UptimeRobot, etc.) | Built-in health checks |
| Total | $25 - $76/mo + significant time | $35 - $60/mo |
For businesses, the reliability and zero-maintenance aspect of managed hosting is the real value. Downtime means missed customer messages and lost revenue. The difference between $35/mo and $25/mo is trivial compared to the cost of your time managing infrastructure — or the cost of a missed customer inquiry.
Ways to Reduce Your Costs
Regardless of how you host, here are practical ways to keep costs down:
- Use a budget model for routine tasks — DeepSeek V3.2 and GPT-4.1 Mini deliver great results for most conversational use cases at a fraction of the cost of frontier models.
- Set system prompts that encourage concise responses — output tokens cost more than input tokens. A prompt like "Be concise and direct" can meaningfully reduce your token usage.
- Disable skills you don't use — web browsing and code execution consume extra tokens for tool calls and results. Only enable what you need.
- Set spending limits on OpenRouter — configure a monthly budget cap so there are no surprises. If the limit is hit, the bot stops responding (better than an unexpected bill).
- Monitor your usage — check your OpenRouter dashboard weekly. If costs spike, investigate whether it's legitimate increased usage or something unexpected.
Frequently Asked Questions
Is OpenClaw itself free?
Yes. OpenClaw is open-source software released under the MIT license. You can download, run, and modify it for free. The costs are for hosting (a VPS or managed platform) and AI model API credits.
Do I need to pay for an API key?
Yes. You need credits with an AI model provider. OpenRouter is the easiest option — it gives you access to all major models through a single API key, with pay-as-you-go pricing.
Can I try before I commit?
On OpenClaw Launch, the Lite plan starts at $3 for the first month. That's enough to deploy an agent and test it thoroughly before deciding if it's worth continuing.
What if I go over my API budget?
OpenRouter lets you set hard spending limits. Once you hit the cap, API calls stop. No surprise charges. You can increase the limit at any time from the OpenRouter dashboard.
Is self-hosting really free with Oracle Cloud?
Oracle offers an always-free tier with Arm-based VMs that are powerful enough to run OpenClaw. The catch: availability is limited and instances can be reclaimed during peak demand. It's a great option for experimentation but not ideal for production use where uptime matters.
The Bottom Line
Running an AI agent with OpenClaw costs anywhere from $4/month (light personal use on a managed platform with a budget model) to $80+/month (heavy business use with premium models). The AI model is the biggest cost driver by far — not the hosting.
For most people, OpenClaw Launch at $3-$20/month plus API costs is the most practical choice: you get a working AI agent in under a minute, with zero infrastructure to manage. Self-hosting makes sense if you have the technical skills and want maximum control — but be honest about the value of your time. The hours you spend configuring Docker, debugging networking issues, and monitoring uptime have a real cost, even if it doesn't show up on an invoice.