AI Unit Economics - 4 min read
LLM Token Unit Economics: Prompt Caching, Rate Limits & AI SaaS Gross Margins
How to model per-user OpenAI, Anthropic, and Gemini API costs, enforce sliding-window token caps, and keep AI subscription gross margins above 75%.
Why flat-rate AI subscriptions silently lose money on power users
Unlike classical SaaS where marginal request cost is near zero, a single heavy user or automated scraper running long-context agent loops can burn USD 60 to USD 150 in LLM tokens inside a USD 29/month plan.
Healthy AI SaaS tiers price against the 90th-percentile active user while enforcing hard per-minute sliding-window rate limits, input character caps, and daily token budgets.
Structure system prompts for 60%+ prompt-cache hit rates
Place static instructions, tool schemas, and reference documents at the exact prefix of your prompt and append dynamic user turns at the tail. Cached input tokens bill at 75% to 90% discounts across major frontier providers, cutting per-turn input spend immediately.