Before you price your AI product, you need to know one number: cost per MAU (monthly active user). Get this wrong and you'll either price yourself out of the market or run out of runway.
The Core Formula
Cost per MAU = (avg_messages_per_user × cost_per_message)
Cost per message =
(input_tokens / 1000) × input_price_per_K +
(output_tokens / 1000) × output_price_per_K
That's it. The complexity comes from estimating the inputs correctly.
Real Benchmarks by Product Type
AI Writing Assistant (Notion AI-style)
Assumptions:
- Model: GPT-4o mini
- Avg messages/user/month: 25
- Input tokens: 1,200 (doc context + instruction)
- Output tokens: 800 (generated text)
Cost per message: (1,200/1000 × $0.00015) + (800/1000 × $0.0006) = $0.00066
Cost per MAU: 25 × $0.00066 = $0.017/user/month
At $10/month subscription, AI costs are 0.17% of revenue. This is a great ratio — the product can afford to be generous with AI usage.
AI Customer Support Bot (Intercom-style)
Assumptions:
- Model: Claude Haiku 4
- Avg tickets resolved/user/month: 8
- Input tokens: 600 (ticket + knowledge base)
- Output tokens: 250 (response)
Cost per message: (600/1000 × $0.0008) + (250/1000 × $0.004) = $0.00148
Cost per MAU: 8 × $0.00148 = $0.012/user/month
For enterprise SaaS at $50+ ARPU, this is negligible. For a $5/month SMB plan, it's still only 0.24% of revenue.
AI Coding Assistant (Cursor-style)
Assumptions:
- Model: GPT-4o
- Avg interactions/user/month: 200 (power users)
- Input tokens: 4,000 (codebase context)
- Output tokens: 600 (suggestion)
Cost per message: (4,000/1000 × $0.0025) + (600/1000 × $0.01) = $0.016
Cost per MAU: 200 × $0.016 = $3.20/user/month
This is the dangerous range. At $20/month, AI costs are 16% of revenue — workable but requires careful optimization (prompt caching, model routing for simple completions).
The Unit Economics Test
Healthy AI product economics:
| Ratio | Benchmark | Action if over |
|---|---|---|
| AI cost / ARPU | < 15% | Switch to cheaper model or add usage limits |
| AI cost / gross margin | < 30% | Revisit pricing or add a usage-based tier |
| AI cost growth vs revenue growth | < 1:1 | Implement per-user caps or caching |
Common Mistakes in Cost Estimation
1. Using average users, not active users Your cost scales with usage, not signups. A user who signs up and never uses the AI feature costs you nothing. Model cost per MAU, not per registered user.
2. Ignoring context window growth As users' conversation history grows, so does your average input token count. Day 1: 500 tokens. Day 30: 3,000 tokens. Model for this — either truncate history, summarize, or factor in the growth curve.
3. Forgetting system prompt overhead A 2,000-token system prompt that you send on every request costs as much as the user's actual message at similar volumes. Enable prompt caching and measure the ratio of static to dynamic tokens.
4. Not accounting for retries and errors Real-world error rates (rate limits, timeouts) add 1–5% to your bill. Build this into estimates.
Modeling for Scale
The most important calculation is what your costs look like at 10× current scale. LLM costs are linear (unlike infrastructure which has some economies of scale). If you're spending $1,000/month today with 1,000 MAU, plan for $10,000/month at 10,000 MAU — unless you optimize.
Use the AI Cost per User Calculator to model your specific numbers, compare models, and stress-test your unit economics before your next pricing decision.
When AI Costs Become a Problem
Warning signs that your AI economics need attention:
- Cost per MAU > 20% of ARPU: Time to add usage-based pricing tiers
- Top 5% of users consuming > 50% of AI costs: Add power-user caps or a higher tier
- AI cost growing faster than revenue: You have a model routing or caching problem
- Gross margin below 50%: AI might be a cost center, not a moat
The goal is to reach a state where AI cost scales slower than revenue — through caching, model optimization, or usage-based pricing that captures value from heavy users.