What Your Product Actually Costs to Run
The API bill is not a one-time cost. It scales with every user action — and most builders discover this at the worst possible moment.
The Bill That Shows Up Uninvited
Here is a scenario that plays out constantly in AI SaaS:
A builder spends three months building a product. They validate it with friends, test it on free API tiers, and launch. Signups trickle in, then accelerate. Month three: 300 users. The API bill arrives: $2,200. The product is priced at $9/month. Revenue: $2,700. After API costs, infrastructure, and their time: they made nothing.
They don't understand how this happened. They thought 300 users at $9 was $2,700/month in the bank.
The missing piece is per-user cost math — and most founders skip it entirely.
Why API Credits Are Different From Every Other Cost
Traditional SaaS infrastructure scales predictably. Your server costs might double when you go from 100 to 1,000 users, but they don't scale 1:1. You get leverage.
AI SaaS doesn't work that way. Every user action that calls an LLM generates a proportional cost. If a user makes 20 requests per day, that's 20 API calls. Multiply by 30 days, multiply by your user count, and you have a number that grows in exact lockstep with your revenue.
This isn't a problem if you've priced for it. It's catastrophic if you haven't.
The Math Walk-Through
Let's make this concrete. One practitioner fact first: every major LLM API prices input and output tokens separately, and output tokens cost several times more than input tokens. GPT-4o, for example, runs about $0.0025 per 1,000 input tokens and $0.01 per 1,000 output tokens at production pricing — a 4× spread. Your input/output mix changes your real cost per request, which is why prompt bloat and verbose responses both show up directly on your bill.
To keep the arithmetic in this track readable, we use a simplified blended rate of $0.005 per 1,000 tokens for the math that follows. It is in the right range for a typical input-heavy request mix, but when you model your own product, always price input and output separately from your provider's current price sheet.
A typical AI SaaS request:
- Input tokens: 2,000 (system prompt + user query + context)
- Output tokens: 500 (generated response)
- Total: 2,500 tokens
- Cost per request (blended rate): 2,500 ÷ 1,000 × $0.005 = $0.0125 per request
A typical active user:
- 20 requests per day
- 30 days per month
- 600 requests per month
- API cost per user: 600 × $0.0125 = $7.50/month
Now scale that:
- 100 users: $750/month in API costs
- 1,000 users: $7,500/month in API costs
- 10,000 users: $75,000/month in API costs
This is only the API layer. Add:
- Hosting (Vercel/Railway): $20–100/month
- Auth service (Supabase): $25/month + overages
- Monitoring (Sentry): $20–50/month
- Your time maintaining prompts: 2–4 hours/month × your hourly rate
The Free Tier Trap
Many builders test on free tiers. OpenAI, Anthropic, and Google all offer free or cheap API access for development. This is genuinely useful for building and validating.
The trap is extrapolating from free tier to production pricing before doing the math.
Development usage is typically:
- Low volume (a few requests per test run)
- Small context windows (simplified prompts)
- Infrequent (you're not running 24/7)
Production usage is:
- Every user action, continuously
- Full production prompts with system context
- Growing over time as more users join
A product that costs pennies per day in development can cost thousands per month at 200 active users. The gap between these two numbers is the gap between "this is profitable" and "I'm paying people to use my product."
The Real Cost of Your Time
Most cost calculations forget the founder's time.
Prompts break. Models change pricing. Edge cases surface that require prompt tuning. A power user finds a weird input pattern that makes your LLM behave unexpectedly. You spend two hours on a Saturday fixing it.
That's a real cost. If you value your time at $100/hour (conservative for a technical founder), 2 hours/month is $200 in invisible cost that never appears on any invoice.
Add this to your cost model. Not because it changes the API math, but because it changes whether the business is worth running.
What to Do With This Knowledge
Before you set a price, calculate your per-user monthly cost at production API pricing. Not dev pricing. Not the free tier. Production.
Then ask: what price gives me a sustainable gross margin after all costs — API, infrastructure, auth, and my time?
If you can't answer that question before you launch, you're guessing. You're hoping that your pricing happens to work out. And hope is not a unit economics strategy.
The next lesson walks through how to use your cost structure to set a price that actually works.
You can dig deeper into the business side of building AI products at jeremyknox.ai.