How Much Does It Cost to Run an AI Agent 24/7? (Real Numbers, 2026)

The question I get asked most often isn't "how does it work" — it's "how much does it cost." Here are the real numbers from a live system that's been running since March 2026.

The Short Answer

Running a lightly loaded AI agent 24/7 on a Mac mini costs $15–40 AUD/month ongoing. Here's the breakdown:

The Mac mini itself is a one-time cost (~$900 AUD for the M4 base model). After that, the main variable is API consumption.

The Biggest Variable: API Costs

Claude Sonnet 4.6 pricing (as of March 2026): $3 USD per million input tokens, $15 per million output tokens.

What does that translate to in practice?

• A single heartbeat check (every 30 min): ~2,000–4,000 tokens → roughly $0.01 AUD
• A cron task that drafts content: ~8,000–15,000 tokens → roughly $0.04–$0.08 AUD
• An overnight autonomous task: ~20,000–50,000 tokens → roughly $0.10–$0.25 AUD
• A long interactive session (30+ back-and-forths, full context): ~100,000+ tokens → $0.50–$2+ AUD

A system running 12 heartbeats/day + 4–6 cron tasks + the occasional long session comes to roughly $3–8 USD/month ($5–12 AUD) in API fees. My first week cost $4 AUD in API fees.

What Spikes Costs

The token consumption that catches people out:

• Long context windows — a session where MEMORY.md, STATUS.md, AGENTS.md and several other files are all in context simultaneously. Each turn re-sends all of that.
• Agent loops — tasks that call tools repeatedly (web scraping, file operations, multi-step automation) can balloon to 50K+ tokens before you notice.
• Large file reads — if your agent regularly reads 10,000-word documents, that cost adds up.
• Vision calls — image analysis costs more per token than text.

What Keeps Costs Down

• Cron isolation — OpenClaw's isolated cron sessions don't carry conversation history. Each cron turn starts fresh. This is the single biggest cost control lever.
• Local models for cheap tasks — I run Qwen 3.5 via Ollama for tasks that don't need Claude-level reasoning. Zero API cost per call.
• Memory architecture — curated MEMORY.md (under 100 lines) vs dumping everything into context. See the memory architecture post for details.
• Targeted heartbeats — heartbeat checks that do one thing (check email, not "do everything") cost far less than broad prompts.

Electricity: Cheaper Than You'd Think

A Mac mini M4 draws about 10–15W at idle, rising to ~35W under load. Running continuously:

• ~12W average × 720 hours/month = 8.6 kWh/month
• Perth electricity rate: ~$0.35 AUD/kWh
• Monthly power cost: ~$3 AUD

With other household devices on the same circuit (router, etc.), I account for ~$5–8 AUD total. Far less than a VPS of equivalent capability, which would run $30–80 AUD/month for an M4-equivalent.

Services: Most Are Free

I've deliberately chosen services with free tiers that cover this scale:

• Cloudflare Pages — free tier handles unlimited static deployments. The business product (aussieclaw.ai) is hosted at zero cost.
• Resend — 3,000 emails/month free. More than enough for product delivery.
• Substack — free to run a newsletter. They take a cut of paid subscriptions (10%), which I consider a distribution fee, not a cost.
• Stripe — 1.7% + $0.30 AUD per transaction. This is a revenue cost, not an operational cost.
• Telegram Bot API — free.

The main non-free service is the domain: $24 AUD/year ($2/month) via Cloudflare Registrar.

Hardware Amortisation

If you're buying a Mac mini specifically for this:

• Mac mini M4 base (16GB RAM, 256GB SSD): ~$899 AUD
• Amortised over 3 years: ~$25 AUD/month

If you already own a Mac that runs 24/7 (or want to repurpose an existing machine), this cost is zero.

Adding hardware amortisation, the true fully-loaded cost for a new buyer is $35–55 AUD/month for the first 3 years, dropping to $12–26 AUD/month after that.

Local Models: The Free Option

If you want to reduce API costs to near-zero, Ollama lets you run open-source models locally:

• Qwen 3.5 (7B) — 6.6GB, runs fine on 16GB Mac mini. Good for summarisation, simple tasks.
• Llama 3.2 (3B) — 2GB, very fast, good for quick routing decisions.
• Gemma 3 (9B) — Strong reasoning for the size, 9GB.

The trade-off: local models are weaker on complex reasoning, multi-step planning, and code quality. The practical setup is a hybrid — local models for cheap, frequent, simple tasks; Claude for anything important.

Real First Week Costs

For transparency, here's what my first 7 days actually cost (March 2026):

• Anthropic API: $4.20 AUD (heavy initial setup sessions, then settled)
• Electricity: ~$1.20 AUD (7/30 of the monthly rate)
• Domain (pro-rated): ~$0.45 AUD
• Services: $0
• Total: $5.85 AUD

First week revenue: $29 AUD (one Guide 01 sale).

That's a $23 AUD net positive on week one. The agent wasn't profitable on day one — it was profitable on day one of its first sale.

Is It Worth It?

The question isn't whether you can afford $15–40/month. The question is whether what you're building with that time is worth it.

A 24/7 agent that handles monitoring, content, scheduling, email triage, and research while you're asleep is replacing what would cost $2,000–5,000/month in human labour. At $40/month, the maths are not close.

The real cost is the setup time — understanding how memory, context, and cron architecture work. That's what the guides are for.

Frequently Asked Questions

How much does it cost to run an AI agent per month?

A lightly loaded AI agent running on a Mac mini costs roughly $15–40 AUD/month. This breaks down as: $4–15 AUD in Anthropic API fees, $5–8 AUD electricity, and $2–5 AUD in minor services (domain, hosting). Heavier use can push API costs above $50/month.

Is it cheaper to run a local AI model instead?

For simple, frequent tasks: yes. Local models like Qwen via Ollama cost zero per call. But local models are significantly weaker on reasoning, coding, and complex instructions. The best setup uses local models for cheap/fast tasks and a cloud model (Claude) for anything that needs quality.

What's the biggest ongoing cost?

Anthropic API fees are the main variable cost. Token consumption spikes during long-context sessions, agent runs with many tool calls, and any task that loops. The fix is using cron jobs rather than long interactive sessions — short, focused tasks cost far less than leaving a session open for hours.

Do I need a Mac mini, or can I use a laptop?

Any Mac works. A MacBook will work while plugged in and with sleep disabled. A Mac mini is preferred for 24/7 because it's designed to run always-on, uses less power, and won't have battery wear from continuous charging. The M4 base model is the recommended starting point.

Can I run this on a Windows PC or Linux server?

OpenClaw is macOS-native. Linux support is on the roadmap. For now, a Mac is required for the full feature set — particularly the Telegram integration, LaunchAgent auto-start, and the browser automation tools.