ai-subscription-unit-economicslisted

# ai-subscription-unit-economics For AI products, **usage limits are a pricing decision**, not a technical afterthought. Design them together or the heaviest 5% of users decide your margin. ## Step 1 — Cost per action, bottom-up For each billable user action, count the tokens honestly: ``` cost_per_action = Σ over model calls: (input_tokens × input_price_per_token) + (output_tokens × output_price_per_token) + (cached_input_tokens × cache_read_price) + non-LLM costs (storage, egress, TTS/image/video, third-party APIs) ``` Where people get it wrong: - **Forgetting retries and failures.** Budget the real retry rate; a failed call still costs input tokens. - **Forgetting the system prompt and context** — often larger than the user's message, and paid on *every* turn. - **Forgetting multi-turn growth.** In a conversation, cost per turn rises as history accumulates. Model the whole session, not one call. - **Assuming a cache hit rate** you haven't measured. Prompt caching is a large lever — verify current pricing and TTL rather than recalling it. - **Agentic loops.** One user action can be N model calls. Count N at the p50 *and* the p95. Get current model prices from the source before computing — do not price from memory. (`claude-api` skill for Claude models.) ## Step 2 — Cost per user ``` cost_per_user_month = actions_per_month × cost_per_action ``` Compute at three points — you're pricing a distribution, not an average: | Cohort | Why it matters