AI Cost Model Template for AI Products

Multiply input tokens by the model's input price per token, add output tokens multiplied by the output price per token. Include any embedding costs for RAG queries. Do not forget to account for retries (failed requests still cost tokens). The [LLM Cost Estimator](/tools) automates this calculation for common models.

When a cheaper model delivers acceptable quality for a given feature. Run your evaluation test suite on the cheaper model. If accuracy drops less than 5% and user satisfaction remains stable, the switch is usually worth it. Start by routing a small percentage of traffic to the cheaper model and comparing metrics side by side.

Set up three defenses: budget alerts at 80% of monthly cap, auto-throttling at 95% that queues requests instead of rejecting them, and a hard spending cap that requires manual approval to override. Also implement request rate limits per user to prevent individual users from driving disproportionate costs.

It depends on the product's value proposition and pricing. AI-native products (where AI is the core value) can sustain 15-25% of revenue on AI costs. Products where AI is a feature among many should target 5-10%. If AI costs exceed 30% of revenue per user, either raise prices, optimize costs, or reconsider the feature. Track the [AI cost efficiency ratio](/glossary/prioritization) alongside revenue metrics.

Committed-use discounts (10-30% savings) make sense when you have at least 3 months of stable usage data and predictable growth. Do not prepay during early product development when usage patterns are volatile. Once you have a stable baseline and forecast, committed pricing can significantly reduce costs.