Lesson 1: The New Economics of Building
Course: The AI-First Startup Playbook
Estimated time: 20 minutes
Difficulty: Intermediate Builder
Everyone says "you can build a startup for nothing now." That's half true. The part that's true: the cost to build has dropped dramatically. You can go from idea to working product in a weekend instead of a quarter. No team required. Infrastructure costs for a small product are genuinely close to zero at launch.
The part that's not true: running an AI-powered product isn't free. API calls cost money. Every user who gets value from your product generates a bill. The economics are different from traditional software in ways that bite founders who don't understand them before they launch.
This lesson gives you the real numbers so you can build the right product at the right price.
What It Costs to Build
The build cost for a solo AI product in 2026 is close to zero if you already code. If you don't, it's low.
Development tools: Cursor or Windsurf costs $20-40/month. Claude or GPT-4 API costs for development are usually under $50/month (you're running short sessions, not high-volume production queries). Total tooling: under $100/month.
Infrastructure at launch:
- Vercel for hosting: free tier handles most early traffic. Pro is $20/month.
- Supabase for database: free tier is generous. Pro is $25/month.
- Railway or Fly.io for backend services: $5-20/month depending on workload.
At launch, before you have significant users, your infrastructure bill is typically $0-45/month. That part of the "build for nothing" claim is accurate.
One-time costs that catch people off guard:
- Domain: $12-15/year
- Email service (Resend, Postmark): free up to a few hundred emails/month, then $20+/month
- Payment processing: Stripe or Lemon Squeezy are free until you charge money, then 2.9% + 30 cents per transaction (Stripe) or similar fees
The AI API Cost Problem
Here's where things get interesting. Every time a user of your product triggers an AI feature, you pay for it. The model doesn't run on your server. You pay per token, per call.
Current rates (approximate, as of early 2026):
GPT-4o: $2.50 per million input tokens, $10 per million output tokens
Claude Sonnet 4: $3 per million input tokens, $15 per million output tokens
Claude Haiku: $0.25 per million input tokens, $1.25 per million output tokens
Gemini 2.0 Flash: $0.075 per million tokens (blended, roughly)
What does a million tokens mean in practice? About 750,000 words of input. A typical GPT-4o API call for a feature like "summarize this document" costs somewhere between $0.001 and $0.05 depending on document length and response size.
That sounds cheap. And for low-volume tools, it is. The math changes at scale.
Example: a document summarization tool
Say your tool takes a 2,000-word document and returns a 200-word summary. With Claude Sonnet:
- Input: ~2,700 tokens (document + system prompt) at $3/million = $0.008
- Output: ~270 tokens at $15/million = $0.004
- Cost per call: ~$0.012
Now say you charge $10/month and users average 200 summaries per month:
- Revenue per user: $10
- AI cost per user: $2.40
- Margin before infrastructure: 76%
That's fine. But if users run 1,000 summaries per month:
- Revenue per user: $10
- AI cost per user: $12
- You're losing $2 per user
This is the trap. It's easy to build a product where users who get the most value from it cost you more than they pay. Free tiers are especially dangerous because heavy users cost you money and pay you nothing.
Modeling Your Product's Costs
Before you launch, you need a unit economics model. It doesn't need to be elaborate - a spreadsheet with honest numbers is enough.
The variables:
Average tokens per user action: Test your actual feature, not your intuition. Run 10-20 sample calls and measure the average token count.
User actions per user per month: How often does a user actually use the feature? Be conservative. Most products have power users and casual users. Model for both.
Cost per token at your model choice: Pick your model, get the current pricing from the provider's pricing page.
Revenue per user per month: Your subscription price, or average revenue per user if you're usage-based.
The formula: (avg tokens per action) x (avg actions per user per month) x (cost per token) = AI cost per user per month.
Compare that to your revenue per user per month. That's your gross margin before infrastructure, payment processing, and support.
If the margin is positive and stays positive as usage grows, you have a sustainable model.
If the margin goes negative at high usage, you need usage limits, usage-based pricing, or a cheaper model.
Model Selection Is an Economics Decision
The choice between models isn't just about capability - it's about cost per query and how that maps to what you can charge.
The practical framework:
Use the cheapest model that produces acceptable output for your use case. Not the best model. The most capable-for-your-needs model.
For many tasks, Claude Haiku or GPT-4o-mini produces output that's 85-90% as good as the flagship models at 10-20% of the cost. If your product is summarizing support tickets, classifying text, extracting structured data from documents, or doing any high-volume, relatively structured task - test the cheaper models first.
Reserve the flagship models (Claude Opus, GPT-4o) for tasks where quality genuinely matters at the output level: creative work, complex reasoning, nuanced writing.
A common pattern: use a cheap model for 90% of operations and a expensive model only for the step where quality is most visible to users.
The $0 to Launch Reality
Here's an honest breakdown of what you actually spend to go from idea to launched product:
Week 1-2 (Building):
- Cursor or AI editor: $40 (month 1)
- Claude API for development calls: $20-50
- Domain: $15
- Vercel free tier: $0
- Supabase free tier: $0
- Total: $75-105
Week 3-4 (Launching):
- Stripe setup: $0 (pay on revenue)
- Email service (Resend): $0 for first 3,000 emails/month
- Any paid marketing (optional): $0-100
- Total added: $0-100
First month running:
- Infrastructure: $0-45 (depending on traffic)
- AI API costs: depends on users (if you have 0 users, $0)
- Total: $0-45
So: the path from idea to launched product in one month, with zero users, costs approximately $75-150 in real money. That's a genuine shift from where software development was 5 years ago.
The cost becomes significant when you have users. Which is the right time for it to become significant.
Where Costs Spike Unexpectedly
A few situations that catch builders off guard:
Agentic features. An agent that makes 15 tool calls per task and reads multiple documents uses 10-50x more tokens than a single-prompt feature. If you build agentic workflows, model the token usage per run carefully before pricing your product.
Long context. Some documents are long. If your product processes PDFs and users upload 50-page documents, your input token count per call is 5-10x what you modeled for a typical document. Add a file size limit or charge more for large documents.
Streaming. Streaming responses don't change token costs, but they do change your infrastructure architecture. A streaming response keeps the connection open longer and requires different handling in your backend. Factor in compute costs for long-running streaming sessions.
Rate limits. When you're on a free or low tier API plan, you'll hit rate limits during usage spikes. Either upgrade your API tier or build queuing into your architecture from the start.
Exercise: Calculate Your Product's Monthly Cost
Pick your product idea. If you don't have one, use this: a tool that takes a job description and generates a tailored cover letter.
Work through this model:
Step 1: Measure token usage. Build the core feature (or estimate based on a prompt you'd use). Run 5-10 sample calls with realistic inputs. Record the token counts from the API response. Calculate the average.
Step 2: Estimate user behavior. How often would a typical user use this feature per month? A cover letter tool: probably 10-30 times while job searching. A daily summarization tool: probably 20-50 times per month.
Step 3: Run the cost math. (avg tokens per call) x (avg calls/month) x (cost per token for your chosen model) = AI cost per user per month.
Step 4: Check the margin. What would you charge per month? Subtract AI cost, subtract $5 for infrastructure share per user (rough estimate). What's left is your gross margin.
Step 5: Model worst case. What if your top 10% of users use it 5x more than average? What does that cost? Can your pricing handle it?
If the margins look good: you have a viable model. If they don't: either pick a cheaper model, add usage limits, or rethink the pricing.
Key Takeaway
The cost to build is near zero. The cost to run scales with usage. Know your token costs per user before you set your prices, or you'll build a product that gets worse for your business as it gets more popular.
What's Next
You know what your product costs to run. Lesson 2 is about making sure enough people will pay for it - finding a market where AI creates a real advantage that people will open their wallets for.