Cost Management for AI APIs

AI API costs have three drivers: input tokens (what you send to the model — system prompts, user messages, context), output tokens (what the model generates — typically 3–5x the cost per token of input), and frequency (how often you call the API). Reducing any of these reduces cost. The most common cost optimisation mistakes: (1) Using an expensive model for simple tasks — Claude Opus for a task Haiku handles equally well. (2) Sending the full system prompt on every request when prompt caching would reduce input token cost by 80–90%. (3) Not batching requests — making 100 individual API calls when one call with 100 items would work. (4) No caching of deterministic responses — if the same question gets the same answer, cache the result at the application layer. For BathCheck's report generation: if a user regenerates the same report 5 times (common during testing), you should cache the first result and serve it from cache. If a user's assessment data hasn't changed, the report won't change either. This single optimisation can reduce Claude API costs by 30–40% in a product with a report preview feature.

Set up cost monitoring before you have a problem, not after. Anthropic's console shows usage by API key — check it weekly initially, then monthly once you understand your patterns. Set up a billing alert at your monthly budget threshold. Cost projection: calculate your cost per user-action (cost per report generated, cost per chat message processed). Then multiply by your projected user volume. If one assessment report costs $0.02 at Claude Sonnet pricing and you process 1,000 reports/month, that's $20/month — trivial. If your system prompt is 4,000 tokens and you have 10,000 API calls per month with no caching, that's 40M prompt tokens = $120/month at Sonnet input pricing — significant but manageable. A practical cost dashboard: add an api_calls table to your Supabase database. Log every Claude API call with: model, input_tokens, output_tokens, cost_calculated, user_id, feature_used. Query this weekly. You'll immediately see which features are most expensive per user and which users are outlier high-consumers (potential abuse or power users who might upgrade to an enterprise plan).