The two line items
Hosting is the boring, predictable part. An agent that sits and waits for messages needs almost nothing: a 1-vCPU, 4 GB VPS (Hostinger KVM 1 is about $7/month; DigitalOcean and Hetzner have equivalents from $5–8) runs OpenClaw, a Discord bot and a couple of cron jobs comfortably. Move to $15–30 when you're running a browser, a vector database or several agents at once. Running it on a spare laptop or a Raspberry Pi is $0 and fine for personal use.
Tokens are where the variance lives. Cost = calls × (input tokens × input price + output tokens × output price), and the input side is usually 80–90% of the bill because every call carries the system prompt, the agent's instructions, tool definitions and whatever context it's holding. Prompt caching cuts the input side by up to 90% on stable prompts, which is why the cache setting above matters more than the model choice for many agents.
Typical monthly costs
| Setup | Calls / month | Tokens (mid-tier model) | Total with VPS |
| Personal assistant, light | < 5,000 | $1–6 | $6–13 |
| Small-business bot | 5,000–10,000 | $15–35 | $25–50 |
| Team automation | 10,000–50,000 | $35–80 | $50–100 |
| Heavy automation | 50,000+ | $80–150+ | $100–200+ |
Ranges assume ~1,000 input / 500 output tokens per call on a Haiku/GPT-mini-class model, per Hostinger's OpenClaw cost analysis (Aug 2026). Frontier models multiply the token line by 3–10×.
What makes the bill explode
- Polling instead of events. An agent that "checks" every 5 minutes makes 8,640 calls a month doing nothing. Use webhooks or a cheap non-LLM check, and only call the model when there's something to do.
- Unbounded memory. Agents that append every interaction to their context grow to 100K+ tokens per call within weeks. Summarize, cap history, or store memory outside the prompt and retrieve only what's relevant.
- Frontier model for every step. Route classification, formatting and retries to Haiku or GPT-5.4 mini; reserve Sonnet/Opus for the reasoning step. A 70/30 split typically cuts token cost by more than half.
- Tool-call loops. One user request can be 5–20 model calls. Count calls, not conversations, and set a per-task step limit.
- No cache hygiene. Changing the system prompt (timestamps, random IDs) on every call invalidates the cache and multiplies input cost by up to 10×.
- Extended thinking on simple steps. Thinking tokens bill as output. Turn effort down for routine calls.
OpenClaw specifically
OpenClaw is free; its cost is entirely the VPS and the model you plug in. A light personal deployment — a few dozen messages a day on a small model — runs under $10 a month all in. Point it at Claude Sonnet or GPT-5.5 with a long memory file and it can climb past $50. The calculator's "personal assistant" preset is the common case; change the model and cache settings to see the swing. Keep an eye on the input tokens per call — that's the memory file — and you'll keep the bill flat.
API key vs subscription
Anything that runs unattended or serves other people should be on an API key with a spend limit set in the console. Claude Code's own scheduled tasks and loops can run inside a Pro or Max plan and draw from your weekly allowance, which is convenient for personal automations but shares the same limits as your interactive work — a loop firing every 10 minutes sends your full context every 10 minutes. See the weekly limit calculator for that math.
Price a single request in detail →
Frequently asked questions
How much does a 24/7 AI agent cost?
$6–20/mo for a light personal agent (VPS + a few dollars of tokens); $25–50 for a busy small-business bot; $100–200+ for heavy automation.
How much does OpenClaw cost?
The software is free. VPS ~$7 plus tokens: under $10/mo light, $30–150 heavy on frontier models.
Cheapest model for agents?
GPT-5.4 mini ($0.75/$4.50) or Claude Haiku 4.5 ($1/$5) for routine steps; escalate to Sonnet/Opus only when needed.
Do I need a VPS?
Only for always-on availability. A home machine works for personal use; a $5–7 VPS is the usual next step.
Can I run it on my Claude Max plan?
Claude Code scheduled tasks, yes, within your weekly limits. Third-party frameworks and anything serving others need an API key.
How do I cap the spend?
Set a workspace spend limit in the API console, cap steps per task, and alert on daily token counts.