Originally published on Alchemic Technology. Read with full formatting →
The number one question people ask before setting up a self-hosted AI assistant is: what will this actually cost me per month? It's a fair question, and the honest answer is better than most people expect. For the VPS alone, you're looking at $4–8/month. Add LLM API usage — which can be as low as $0 if you lean on free tiers — and most people running a personal assistant with moderate usage end up spending $5–15/month total. Compare that to a single ChatGPT Plus subscription at $20/month and the math starts looking very interesting.
This guide breaks down every cost component with real numbers: VPS options and what each gives you, LLM pricing by model with honest per-message estimates, the full $0/month free-tier path, and the recommended setup most people should start with. No vague ranges — just actual prices as of early 2026.
The Two Cost Buckets
Running a self-hosted AI assistant has exactly two cost inputs:
- **Server (VPS)** — The machine that runs OpenClaw 24/7. This is a fixed monthly cost regardless of how much you use it.
- **LLM API calls** — What you pay the AI model provider each time you send a message. This scales with usage.
That's it. No other mandatory costs. Domain name is optional. Bandwidth is almost always included. Backups are optional add-ons. Let's price each one out.
VPS Options: What Each Gets You
OpenClaw runs comfortably on any Linux VPS with 1GB+ RAM and 1+ vCPU. Here are the realistic options across the major providers, sorted by monthly cost:
| Provider / Plan | vCPU | RAM | Storage | Price/mo | Notes |
|---|---|---|---|---|---|
| Oracle Cloud Free Tier | 2 | 1 GB | 50 GB | $0 | Always Free — but limited regional availability |
| Hetzner CX22 ⭐ Recommended | 2 | 4 GB | 40 GB SSD | ~$4.15 | Best specs-per-dollar. EU/US locations. Excellent reliability. |
| Hostinger KVM 2 | 2 | 8 GB | 100 GB NVMe | ~$5.99 | Strong value if you want more RAM headroom |
| DigitalOcean Basic Droplet | 1 | 1 GB | 25 GB SSD | $6.00 | Solid reliability, good docs, US/EU/Asia regions |
| Vultr Cloud Compute | 1 | 1 GB | 25 GB SSD | $6.00 | Comparable to DigitalOcean; 30+ global locations |
Why Hetzner CX22 is the recommended pick: At ~$4.15/month, you get 4GB RAM and 2 vCPUs — four times the RAM of the $6 DigitalOcean or Vultr options at a lower price. This matters for OpenClaw, which benefits from having room for the Node.js process, any MCP servers you run locally, and occasional memory spikes during complex agent tasks. Hetzner's reliability is excellent and their Falkenstein (EU) and Ashburn (US) datacenters offer good latency for most users.
LLM API Costs: Model-by-Model Breakdown
This is where self-hosting gets interesting. Unlike a subscription service that bundles one model at a fixed price, you pay per token — and you can mix and match models to optimize cost. Here's what each major model actually costs as of March 2026, pulled from official provider pricing pages:
Anthropic (Claude)
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Best For |
|---|---|---|---|
| Claude Haiku 3.5 ⭐ Best value | $0.80 | $4.00 | Daily tasks, quick queries — dirt cheap at scale |
| Claude Haiku 4.5 | $1.00 | $5.00 | Newer Haiku — slightly smarter, slightly pricier |
| Claude Sonnet 4.6 | $3.00 | $15.00 | Complex reasoning, coding, nuanced writing |
| Claude Opus 4.6 | $5.00 | $25.00 | Hardest tasks — heavy cost, reserve for real complexity |
OpenAI
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Best For |
|---|---|---|---|
| GPT-4.1-nano ⭐ Cheapest OpenAI | $0.10 | $0.40 | Ultra-cheap routing, classification, high-volume tasks |
| GPT-4o-mini | $0.15 | $0.60 | Lightweight tasks, solid quality at low cost |
| GPT-5-mini | $0.25 | $2.00 | GPT-5 quality at budget pricing — strong everyday model |
| GPT-4.1-mini | $0.40 | $1.60 | Balanced cost/quality, good for agents |
| GPT-4.1 | $2.00 | $8.00 | Current GPT-4 flagship — great multimodal and tool use |
| GPT-4o | $2.50 | $10.00 | Vision, voice, legacy integrations |
| GPT-5 / GPT-5.1 | $1.25 | $10.00 | Next-gen flagship — surprisingly cost-competitive |
| GPT-5.4 | $2.50 | $15.00 | Top-tier capability — the current best from OpenAI |
Google (Gemini)
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Free Tier | Best For |
|---|---|---|---|---|
| Gemini 2.5 Flash ⭐ Free tier pick | $0.15 | $0.60 | ✅ 15 RPM, generous daily limit | High-volume personal use — best free option |
| Gemini 2.5 Flash-Lite | $0.10 | $0.40 | ✅ Free tier available | Ultra-cheap routing and classification tasks |
| Gemini 2.5 Pro | $4.00 | $20.00 | No | Top-tier reasoning — competes with Claude Sonnet |
MiniMax
| Model / Plan | Input (per 1M tokens) | Output (per 1M tokens) | Notes |
|---|---|---|---|
| M2.5 Standard ⭐ | $0.15 | $1.20 | ~50 TPS, automatic caching, no config needed |
| M2.5 Lightning | $0.30 | $2.40 | ~100 TPS — same performance, twice the speed |
**💡 MiniMax Coding Plan (flat-rate subscription):** MiniMax offers a **Coding Plan** — a fixed monthly subscription that grants a set number of prompts for AI coding tools (Cursor, Windsurf, etc.) using M2.5 / M2.5-highspeed. Three tiers: **Starter, Plus, Max**. MiniMax claims it runs at ~1/10th the cost of equivalent Claude plans. If you do heavy AI-assisted coding, this is worth checking out at [platform.minimax.io/subscribe/coding-plan](https://platform.minimax.io/subscribe/coding-plan). The Coding Plan API key is separate from the standard pay-as-you-go key.
**📌 Note on OpenClaw + MiniMax:** MiniMax M2.5 is available via the OpenClaw portal integration — often at no additional cost for background tasks and cron jobs. If you're running OpenClaw, M2.5 is an excellent zero-to-low-cost option for all your scheduled automation work.
What This Means Per Message
Prices per million tokens sound abstract. Let's ground them in reality. A typical back-and-forth message exchange (your message + the AI response) uses roughly 500–2,000 tokens total, depending on complexity.
- **Claude Haiku 3.5:** ~$0.001–0.004 per exchange. Send 500 messages/month and pay about $1. This is the go-to daily driver.
- **Claude Sonnet 4.6:** ~$0.01–0.03 per exchange. Send 200 detailed messages and pay about $4–6. Reserve it for tasks that need real reasoning power.
- **Gemini 2.5 Flash (free tier):** $0 up to the daily rate limit — hundreds of messages daily at zero cost. Rate limits (15 RPM) apply but rarely matter for personal use.
- **GPT-4o-mini:** ~$0.0003–0.001 per exchange — cheaper than Haiku for pure throughput. Good if you're heavily OpenAI-tooled.
- **MiniMax M2.5:** ~$0.001–0.003 per exchange at market rates — free via OpenClaw's provider integration for background/cron usage.
Realistic monthly API cost for moderate personal use: Most people running a personal assistant — asking questions, running daily briefings, doing research tasks — spend $2–8/month on API calls. Heavy users who run multiple sub-agents and complex automation pipelines might hit $15–20/month. Very light users can stay under $1/month by leaning on free tiers.
Path A: The $0/Month Free Tier Setup
Yes, you can run a self-hosted AI assistant for genuinely zero dollars per month. Here's the exact combination:
The Free Stack
- **Server:** Oracle Cloud Always Free — 2 vCPU, 1GB RAM, 50GB storage, $0/month
- **Primary model:** Gemini 2.5 Flash free tier — up to 1M tokens/day, no credit card required
- **Background tasks / crons:** MiniMax M2.5 — free tier with 200K context window
- **Total monthly cost: $0**
This works. OpenClaw runs fine on Oracle's free ARM instances, and Gemini 2.5 Flash's free tier is genuinely generous for personal use.
The honest caveats:
- **Oracle regional availability is spotty.** The Always Free ARM instances frequently show "capacity unavailable" in popular regions. You may need to try multiple regions or wait for availability. Once you have it, you keep it — but provisioning can take patience.
- **Gemini Flash rate limits.** The free tier has per-minute rate limits that can slow down complex multi-step tasks. For casual daily use, you likely won't hit them. For heavy automation pipelines, you will.
- **No performance headroom.** 1GB RAM is the minimum. OpenClaw runs, but you won't have much room for additional local MCP servers or heavy concurrent tasks.
If you can get Oracle provisioned and your usage fits within free-tier limits, this is a legitimate setup. Just go in with expectations calibrated to the constraints.
Path B: The Recommended Setup ($5–8/Month)
This is the setup the OpenClaw Field Guide is built around. It's what most people should start with.
The Recommended Stack
- **Server:** Hetzner CX22 — 2 vCPU, 4GB RAM, 40GB SSD — **~$4.15/month**
- **Primary model:** Claude Haiku 3.5 for everyday queries — ~$1–3/month at moderate usage
- **Power tasks:** Claude Sonnet 4.6 when you need it — occasional use adds $1–3/month
- **Background tasks / crons:** MiniMax M2.5 — free, handles all scheduled work
- **Total monthly cost: ~$5–8/month**
Why this combination works well:
- Hetzner's 4GB RAM gives you real headroom. You can run OpenClaw, a few local MCP servers, and still have memory available for spikes.
- Claude Haiku handles 80–90% of daily tasks at low cost. Quick questions, calendar summaries, web searches, task management — Haiku does all of this well and cheaply.
- Claude Sonnet steps in when you need better reasoning — code review, complex writing, multi-step analysis. Using it selectively keeps costs predictable.
- MiniMax at zero cost for all the background automation means your cron jobs and heartbeat checks don't add to the bill.
The result: a personal AI assistant that runs 24/7, answers messages on Telegram or Discord, handles scheduled tasks automatically, and can route complex requests to a smarter model when needed — for about the cost of two cups of coffee a month.
Hidden Costs to Know About
The VPS + API breakdown covers 95% of the actual cost. But there are a few optional line items worth knowing before you start:
- **Domain name:** Optional, but nice to have for webhooks and a clean URL for your agent's web interface. Typical cost: ~$12/year (~$1/month) through Namecheap or Cloudflare. Not required to run OpenClaw.
- **Bandwidth:** Almost always included in VPS plans at the entry level. Hetzner includes 20TB/month on the CX22. You won't hit it running a personal assistant.
- **Automated backups:** Optional add-on most VPS providers offer. Typically $1–2/month for daily snapshots. Recommended once you've built out your workspace configuration and have files you'd miss.
- **SSL certificate:** Free via Let's Encrypt. No cost here.
Even if you add a domain and backups, you're at about $7–11/month all-in. Still well below a single SaaS subscription.
Cost vs. Subscription Services
Here's the full picture — subscription services vs. pay-as-you-go API access via OpenClaw:
What the subscriptions actually cost (2026)
| Service | Monthly Cost | What You Get | Catch |
|---|---|---|---|
| ChatGPT Free | $0 | Limited GPT-5.3 access, slow image gen, basic memory | Heavily rate-limited; no advanced reasoning |
| ChatGPT Go | ~$8–10/month | Expanded GPT-5.3 messages and uploads | New tier — limited feature set vs Plus |
| ChatGPT Plus | $20/month | Advanced reasoning, Codex agent, Sora, deep research | Usage caps; locked to OpenAI platform only |
| ChatGPT Pro | $200/month | GPT-5.4 Pro (unrestricted), unlimited uploads, max deep research | Only makes sense for heavy power users |
| Claude Free | $0 | Access to Claude models, web search, MCP connectors | Rate-limited; no Projects or Research |
| Claude Pro | $20/month ($17 annual) | More usage, Projects, Research, Claude Code & Cowork | Usage caps apply; no API access or automation |
| Claude Max | $100–$200/month | 5x–20x more usage than Pro, priority access | Expensive — only for extreme daily Claude usage |
| Google AI Pro (Gemini) | $19.99/month | Gemini Advanced, 2TB Drive, expanded AI features | Locked to Google ecosystem; not programmable |
| OpenClaw on Hetzner CX22 | ~$5–8/month total | Claude + GPT-5 + Gemini + MiniMax — all providers at once | None — you own the stack |
**🔑 The key insight:** Subscriptions give you fixed-cost access to *one provider's* chat interface. API access via OpenClaw gives you pay-as-you-go access to *every provider simultaneously* — often cheaper for moderate use, and infinitely more powerful because you can automate, schedule, and route between models.
When subscriptions still make sense
To be fair: subscriptions aren't always the wrong choice. They make sense when:
- **You're a heavy daily ChatGPT user** — if you're sending 100+ messages/day, Claude Pro or ChatGPT Plus is likely cheaper than equivalent API usage
- **You need the consumer UX** — mobile apps, voice mode, native integrations like Claude for Excel/PowerPoint
- **You don't want to manage infrastructure** — subscriptions require zero setup or maintenance
- **You want Sora, DALL-E, or image gen included** — ChatGPT Plus bundles these; via API they're separate line items
The case for self-hosting isn't "subscriptions are always bad." It's that self-hosting with OpenClaw gives you more models, more control, and comparable or lower cost — especially once you start automating things subscriptions can't do at all.
The math for most users: ChatGPT Plus alone is $20/month. Claude Pro is another $20. That's $40/month for two providers with usage caps. OpenClaw on Hetzner at $5–8/month gives you both providers plus Gemini and MiniMax, no caps, with full automation. That's the trade-off.
The Bottom Line
A fully functional, always-on personal AI assistant costs $5–8/month with the recommended setup — or $0/month if you're patient with Oracle's free tier and work within Gemini's rate limits. Either way, you're spending less than any single AI subscription service while getting more flexibility, more model options, and full control over your data and automation.
Where to Go From Here
Knowing the cost is step one. The next step is the actual setup: provisioning the VPS, installing OpenClaw, connecting your first channel, and configuring the workspace files that give your agent its memory and personality. That's what the OpenClaw Field Guide covers in detail — from first SSH connection to a production-ready agent with multi-model routing, sub-agent teams, and scheduled automation.
Ready to Build Your Personal AI Stack?
The OpenClaw Field Guide is 58 pages across 14 chapters covering everything from initial VPS setup to advanced multi-agent automation. It's the complete reference for getting from "installed" to "indispensable" — built around the Hetzner + Claude Haiku setup described in this post.
[Get the OpenClaw Field Guide →](https://guide.alchemictechnology.com)
Want a complete setup guide? The OpenClaw Field Guide covers VPS provisioning, channel setup, model config, skills, and automation in one place.
Top comments (0)