What this actually costs once people use it
You haven't launched yet because you're picturing a $2,000 bill from a service you forgot you connected. That fear is reasonable. Most "surprise" AI app bills come from two or three predictable places, not from mystery charges. Here's what actually costs money, when it starts costing money, and how to put a ceiling on it before you go live.
The real line items
Most small AI-built apps run on some combination of these. Free tiers are generous enough that a lot of people never leave them.
| Service | Free tier (typical) | First paid tier (typical) | What triggers the jump |
|---|---|---|---|
| Hosting (Vercel, Netlify, Render) | 100GB bandwidth/month, unlimited small sites | $19–20/month | High traffic, background jobs, or team seats |
| Database (Supabase, Neon, PlanetScale) | 500MB–1GB storage, limited compute hours | $25/month | Your data grows past the storage cap or you need always-on compute |
| Auth (Clerk, Supabase Auth, Auth0) | Usually free up to several thousand monthly active users | $25/month and up | Active user count, not signups. Someone who logs in counts, someone who never returns doesn't |
| AI API calls (OpenAI, Anthropic, etc.) | No real free tier once you're past trial credit | Pay per token, no flat rate | Every single message sent and received. This is the one people underestimate |
| Email sending (Resend, Postmark) | 100–3,000 emails/month free | $15–20/month | Sending password resets, notifications, or marketing at volume |
| File storage/CDN (S3, Cloudflare R2) | Small free allowance, often 10GB | Pennies per GB after that | Users uploading images or files regularly |
| Domain | N/A | $10–15/year | Fixed, doesn't scale with usage |
Why AI tokens are the one that surprises people
Hosting and databases usually have a flat monthly price once you cross the free line. AI calls don't work that way. You pay per token, every time, for every request. A chat feature that feels free while you're testing it can add up fast once real users start typing longer messages or asking the AI to read documents.
A rough way to think about it: a short back-and-forth conversation might cost a fraction of a cent. A feature that summarizes long documents or runs on every page load can cost 10 to 50 times more per use. The fix isn't to avoid AI features. It's to know the price per call before you ship the feature, not after the bill arrives.
Where to set hard caps and alerts
- AI provider (OpenAI, Anthropic): Set a hard monthly spend limit in the billing dashboard, not just an alert. Alerts tell you after you've spent money. Hard limits stop the spending.
- Hosting: Set a usage alert at 80% of your free tier bandwidth. Most platforms have this under billing or usage settings.
- Database: Set an alert on storage size and on compute hours if your provider bills separately for those.
- Email: Set an alert at your monthly send limit so a bug that loops and sends 10,000 emails doesn't cost you a tier jump overnight.
- Your bank or card: Use a virtual card with its own spend limit for anything connected to AI billing. This is the simplest backstop if you forget to set a cap somewhere else.
Ask your AI to check this before it builds anything
The one-line answer
For a small app with a few hundred users, checking in occasionally rather than constantly, expect $0 to $25 a month total, mostly free tier, with your AI API calls being the only part likely to move that number if a feature gets used a lot more than you expected.
Plyo runs a health check next to your project so this isn't a guessing game as you grow.
Plyo is the home for AI-built projects: every version saved, one click to undo, one click to publish, plus a safety check, visitor counts and backups. Free to start: 3 projects, no card needed.
Give your project a home