When I tell people I run an entire business on AI agents for about $80/month, the first question is always: "What exactly does that $80 cover?"
It's a fair question. The internet is full of breathless claims about AI cost savings, but very few people publish actual numbers. So here's mine — real costs, real use cases, broken down by tier so you can find where you fit.
Short answer: A solo operator running 3–5 agents spends $40–80/month. A small business with 8–12 agents covering real workloads runs $150–400/month. Full business automation with 15+ agents tops out around $300–800/month — still a fraction of equivalent human labor.
The Two Cost Buckets
AI agent costs come from two places:
- API costs — what you pay per token to the model provider (OpenAI, Anthropic, Google, etc.)
- Infrastructure costs — hosting, scheduling, monitoring, integrations
Most people obsess over model costs and underestimate infrastructure, or vice versa. Both matter. Let's break them down.
Model Costs: What You Actually Pay Per Agent
Model pricing is per-token. But what does that mean in practice? Here's the honest translation:
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Est. cost/agent/mo | Best for |
|---|---|---|---|---|
| GPT-4o mini | $0.15 | $0.60 | $3–8 | High-volume routine tasks |
| Claude Haiku 3.5 | $0.80 | $4.00 | $5–12 | Structured output, fast responses |
| GPT-4o | $2.50 | $10.00 | $15–40 | Complex reasoning, ambiguous input |
| Claude Sonnet 4 | $3.00 | $15.00 | $20–60 | Long-form writing, nuanced decisions |
| Claude Opus 4 | $15.00 | $75.00 | $80–200+ | CEO-level judgment, complex orchestration |
The routing rule that cuts my costs by 60%: 90% of agent tasks are routine — summarize, classify, draft a response. GPT-4o mini handles these at $3–8/agent/month. Only 10% of tasks genuinely need a smarter model. Route accordingly and your costs drop dramatically.
The agents that cost real money are the ones with large context windows (lots of history passed every run) and high output volume (generating long documents repeatedly). Know which agents those are before you build.
Infrastructure Costs: The Underestimated Half
Model costs are visible on your API dashboard. Infrastructure costs are what sneak up on you.
| Component | What it does | Cost range/mo |
|---|---|---|
| Hosting / compute | Runs your cron jobs and scripts | $0–20 |
| Agent runtime | OpenClaw, LangGraph, CrewAI, etc. | $0–29 |
| Email service | Newsletters, transactional emails | $0–15 |
| Storage (memory files) | Agent memory persistence | $0–5 |
| Monitoring / alerting | Know when agents fail or go rogue | $0–20 |
| Integrations (CRM, calendar, etc.) | Data in/out of your existing tools | $0–50 |
If you run agents on your own machine or a cheap VPS, infrastructure can be nearly $0 outside model costs. If you're using managed platforms, add $30–80/month.
Real-World Cost Tiers
Tier 1: Solo Operator / Freelancer
- 3–5 agents running (email triage, weekly report, social drafts, support FAQ)
- Mostly GPT-4o mini / Claude Haiku
- Self-hosted or lightweight VPS
- Replaces ~10 hrs/week of manual work
Tier 2: Small Business / Agency
- 8–12 agents (plus content generation, lead research, client reporting, CRM updates)
- Mix of cheap models for routine + mid-tier for client-facing output
- Integration costs start adding up (Zapier/n8n, CRM webhooks)
- Replaces ~30 hrs/week of manual or VA work
Tier 3: Full Business Automation
- 15+ agents across operations, marketing, support, finance, content
- CEO-level reasoning agents (Claude Opus, GPT-4o) running the expensive decisions
- Real infrastructure: dedicated hosting, monitoring, redundancy
- Equivalent to 2–3 FTE in output volume
Context on "Tier 3": $800/month sounds like a lot until you compare it to the cost of 2–3 employees. A single US-based employee costs $50K–80K/year all-in. $800/month is $9,600/year — roughly 12–15% of one salary. The math works if you're replacing real labor.
Where People Overspend (and How to Fix It)
After talking to dozens of operators, the same cost mistakes appear over and over:
1. Wrong model for the task
Running GPT-4o on every task is like using a sports car to haul groceries. 70% of agent tasks are classification, summarization, and templated output — GPT-4o mini does these just as well for 15× less money.
2. Context bloat
Passing the full history of every conversation into every prompt. Most agents need the last 5–10 exchanges plus their core instructions — not the last 300 messages. Compressed context saves 30–50% on input tokens for long-running agents.
3. Loop inefficiency
Agents that check for updates every minute when daily would suffice. A heartbeat agent checking email every 60 seconds makes 43,000 API calls per month — most of them returning "nothing new." Cron it to every 30 minutes and you cut calls by 96%.
4. No cost monitoring
You can't optimize what you don't measure. Set spend alerts on your API dashboards. A runaway agent (infinite loop, misbehaving retry) can burn through $50–200 in hours without one.
Want the exact cost optimization configs?
The Library includes a full guide on multi-model routing — the exact decision tree I use to route tasks to cheap vs. expensive models, plus the cron scheduling patterns that cut my API calls by 60%. 68+ production-tested guides, from $9/month.
See What's in the Library →The ROI Question: What Does This Actually Replace?
Cost without context is meaningless. Here's what $80/month buys in real output:
- Email triage and reply drafts: 5–10 hrs/week recovered
- Weekly business report: 2–3 hrs/week recovered
- Social media drafts: 3–5 hrs/week recovered
- Support FAQ responses: 4–8 hrs/week recovered
That's 14–26 hours/week of labor at $0/hr marginal cost after setup. At even a conservative $30/hr value on your time, $80/month buys back $1,680–3,120/month of capacity.
That's the math. It's not complicated. The hard part is the setup — getting the agents configured right so they actually work reliably and don't create more cleanup work than they save.
The Setup Cost: Don't Forget This One
Monthly cost is easy to project. Setup cost is where people get surprised.
Plan for 8–20 hours of setup time to build a functional 5-agent stack. This includes writing the identity files, connecting integrations, testing edge cases, and tuning each agent until it behaves correctly. It's a one-time investment — but it's real.
The shortcuts that save the most time: starting from tested configs (not from scratch), knowing which failure modes to design around from the beginning, and having clear escalation rules so agents don't spiral when they hit the unexpected.
The Library exists specifically to cut this setup time. Every guide is a config you can adapt, not a concept you have to implement from zero.
Quick Reference
- Solo operator (3–5 agents): $40–80/month
- Small business (8–12 agents): $150–400/month
- Full stack (15+ agents): $300–800/month
- Model to start with: GPT-4o mini (cheap, reliable for 90% of tasks)
- Biggest cost trap: Wrong model for the task + no context compression
- Setup time to budget: 8–20 hours for a working 5-agent stack
Questions about your specific setup? The Workshop tier includes direct Q&A — post your use case and I'll give you a specific answer: askpatrick.co/get-access.