How I Run an AI Business for $120/Month

My total operating cost last month: $117.83. My revenue: $2,940. That's a 96% net margin.

Not hypothetical. Not projected. Actual numbers from actual Stripe and API dashboards. Here's where every dollar went.

The Full Breakdown

Read that bottom line again. The AI that runs my entire content pipeline, analytics, customer comms, and product delivery costs $4.83 per month. The X API tier is 20x more expensive than the intelligence layer.

Why It's So Cheap: Model Routing

The single biggest cost-saving decision I made: stop sending everything to the same model.

Before model routing, I was spending $30–$40/month on API calls. Everything went to Sonnet because it was "good enough" for everything. It was — but so is a Ferrari for grocery runs. You can. You shouldn't.

Here's how tasks actually route now:

# model_router.py — task-based routing
ROUTES = {
    # Tier 1: Haiku ($0.25/$1.25 per 1M tokens) — 73% of all calls
    "tweet_classify":     "haiku",    # $0.0003/call
    "format_check":       "haiku",    # $0.0002/call  
    "engagement_triage":  "haiku",    # $0.0004/call
    "email_categorize":   "haiku",    # $0.0003/call
    "queue_deduplicate":  "haiku",    # $0.0003/call

    # Tier 2: Sonnet ($3/$15 per 1M tokens) — 25% of calls
    "content_draft":      "sonnet",   # $0.003/call
    "weekly_report":      "sonnet",   # $0.04/call
    "email_draft":        "sonnet",   # $0.003/call
    "analytics_summary":  "sonnet",   # $0.002/call

    # Tier 3: Opus ($15/$75 per 1M tokens) — 2% of calls
    "architecture":       "opus",     # $0.02/call
    "playbook_chapter":   "opus",     # $0.08/call
    "voice_audit":        "opus",     # $0.04/call
}

73% of calls go to Haiku. Most "AI work" is classification, formatting, and pattern matching — tasks where the cheapest model performs identically to the most expensive one. You're not getting better tweet classification from Opus. You're just getting a bigger bill.

The Daily Cost Breakdown

Here's what a typical day costs in AI:

Eight cents a day. Running an entire business operation. The dashboard generation and watchdog scripts use zero AI — they're pure Python reading JSON files and doing math. No LLM call needed to check if a number went up or down.

The Scripts That Cost $0

This is the part people miss. The majority of my automation doesn't use AI at all:

• poster.py — reads tweet from queue, posts via API. Zero AI. $0.
• dashboard_gen.py — pulls metrics from Twitter API, Stripe API, file system. Writes JSON. Zero AI. $0.
• watchdog.py — checks file freshness, queue depth, log recency. Zero AI. $0.
• checkout_test.py — hits Stripe payment links, verifies HTTP 200. Zero AI. $0.
• revenue.py — queries Stripe, logs daily revenue. Zero AI. $0.
• queue_manager.py — rotates content, expires timely posts, deduplicates. Zero AI. $0.

Six of my most critical scripts don't make a single API call to any AI model. They're 50–200 lines of Python doing file I/O and HTTP requests. People overcomplicate this. Not every automation needs intelligence. Most need reliability.

Where the X API $100 Goes

The elephant in the room: 80% of my costs are the X API Basic tier. Is it worth it?

At my current revenue ($2,940/mo), that $100 is a 3.4% cost. The X API gives me:

• 10,000 post reads/month (timeline monitoring, engagement tracking)
• Unlimited posts (I use ~150/month across tweets and replies)
• Search access (finding conversations to join)
• Engagement metrics (what's working, what's not)

Could I do this on the free tier? Technically yes — free tier allows 1,500 posts/month. But no read access means no engagement automation, no analytics, no reply monitoring. That's not an operator. That's a broadcast antenna.

The $100 is my distribution cost. My "ad spend." Most businesses spend 20-30% on customer acquisition. I spend 3.4%.

What I Don't Pay For

Things conspicuously absent from my stack:

• No scheduling tool ($25–$50/mo saved) — cron + poster.py does the same thing
• No analytics SaaS ($15–$50/mo saved) — Python scripts + JSON files
• No email marketing platform ($30–$100/mo saved) — API-first email + scripts
• No automation platform ($30–$70/mo saved) — Python scripts, not Zapier
• No project management tool ($10–$25/mo saved) — markdown files
• No CRM ($25–$80/mo saved) — JSON + scripts

Conservative total saved: $135–$375/month in SaaS I replaced with scripts. The scripts took a weekend to write. They'll run forever.

The Margin Math at Scale

Here's where this gets interesting. My costs scale sublinearly — revenue goes up, costs barely move:

AI costs barely double when revenue 5x's because you're doing the same operations — just to more people. Posting 5 tweets costs the same whether 100 or 10,000 people read them. Generating a weekly report costs $0.04 whether revenue is $500 or $25,000.

The only costs that scale with revenue: Stripe fees (fixed 2.9%), email volume (pennies per send), and eventually customer support volume. None of those are AI costs.

How to Get Under $10/Month in AI Costs

Five rules:

1. Route aggressively. If a task doesn't require creativity or judgment, it goes to the cheapest model. Period.
2. Batch, don't stream. Generate 10 tweets in one API call ($0.02), not 10 calls ($0.03 each = $0.30). 15x cheaper.
3. Skip AI where Python works. File operations, HTTP checks, arithmetic, formatting — no LLM needed.
4. Cache everything. Same analytics query twice? Read from yesterday's JSON. Don't recompute.
5. Measure per-task costs. You can't optimize what you don't track. Log every API call with its cost.

# Add this to every script that calls an AI model
import json, time

def log_cost(task: str, model: str, input_tokens: int, output_tokens: int):
    costs = {"haiku": (0.25, 1.25), "sonnet": (3, 15), "opus": (15, 75)}
    inp_cost = input_tokens / 1_000_000 * costs[model][0]
    out_cost = output_tokens / 1_000_000 * costs[model][1]
    total = inp_cost + out_cost
    
    with open("logs/costs.jsonl", "a") as f:
        f.write(json.dumps({
            "task": task, "model": model,
            "cost": round(total, 6),
            "timestamp": time.time()
        }) + "\n")

Track every cent. When your monthly AI bill hits $10, you'll know exactly which tasks to route cheaper.

The Bottom Line

Running an AI business doesn't require a big budget. It requires the right architecture. Model routing alone saves 85% on AI costs. Replacing SaaS with scripts saves another $100–$300/month. And the margin math only gets better as revenue grows.

$120/month. $2,940 in revenue. 96% margin. That's not a side project. That's a business.