Inference Cost
How much it costs to run an LLM inference call.
Inference cost is the per-call expense of using an AI model. For frontier APIs, it's typically priced per million input tokens and per million output tokens, with output usually 3-5x more expensive than input.
For operators building AI products, inference cost is the line that determines unit economics. A naive agent design that uses 100K tokens per user session may be unprofitable; the same agent with prompt caching, model routing (cheap models for easy steps), and bounded loops may be highly profitable.
Key cost levers in 2026: prompt caching (cuts cost for repeated prefixes by 50-90%), model selection (a 10x cost gap between fastest and smartest tiers), batch processing (cheaper if latency-tolerant), and avoiding accidental loops in agent code. Most teams who say "AI is too expensive for us" haven't deeply optimized any of these.