How AI Agent Costs Compound: A Token-by-Token Breakdown and 5 Ways to Save

A 20-turn AI agent run processing roughly 59,000 tokens of content can be billed for over 656,000 input tokens, because stateless model APIs resend the entire conversation history on every step. Using Claude Sonnet 5 as a benchmark at Anthropic's published rates, a 20-turn research agent costs approximately $1.37 without caching, with input tokens — not output — driving over 96% of that bill. Costs scale roughly with the square of the number of turns, meaning doubling from 20 to 40 turns nearly quadruples the expense, driven primarily by accumulating tool results carried forward each turn. Prompt caching can cut the 20-turn cost by 76% to around $0.33, though its effectiveness depends on stable context ordering and fast response times to keep cached data alive. The analysis concludes that understanding token accumulation mechanics — not just per-token rates — is becoming a critical skill for anyone deploying AI agents in production.