Cache Read

A cache read is when an AI agent pulls back information it has already processed and stored from its temporary memory. Think of it as the agent quickly glancing at its notes from five minutes ago. This isn't a fresh input, it is a recall, and it is how agents maintain continuity across a conversation or task without re-ingesting everything from scratch.

This is not a free operation, despite its low cost per token. It is also not the same as a cache hit in traditional computing, where data is retrieved from a fast memory layer to avoid a slower, more expensive computation entirely. For AI agents, a cache read is a specific, cheap billing tier for re-accessing context, not avoiding the token count altogether. You still pay, just significantly less.

The common confusion is that if an agent "remembers" something, it must be free, or at least negligible. That is a fantasy. Every single time the agent needs to reference its system prompt, its defined tools, or the ongoing transcript, it performs a cache read. This constant re-reading is fundamental to how agents operate, ensuring they stay grounded in the current task and do not hallucinate.

Consider Anthropic's Claude Code API. Their documentation, and our own measurements, show a cache read priced at just 0.1 times the base input rate. This is incredibly cheap per token, making it the most cost-effective way for an agent to maintain its working memory.

However, its sheer volume makes it the biggest line item on many invoices. In our three-month study of Claude Code usage, cache reads accounted for a staggering 48% of the total bill. This was for a single workstation running specific design-related tasks, involving 157,670 deduplicated API responses. It proves that even the cheapest per-token operation can dominate costs when it happens hundreds of thousands of times.

You should always aim for cache reads over more expensive alternatives like cache writes or uncached input. The "when not to use" is really "when you fail to get one," which means you either let your session expire or you are working with an agent that does not support caching.

To maximize cache reads, keep your agent sessions active. If your agent idles past its configured time-to-live, typically five minutes for Anthropic models, the next interaction will trigger a more expensive cache-write instead of a cheap read. This means a quick coffee break can literally turn a cheap memory recall into a costly storage event.

The tradeoff is between immediate cost and persistent context. You want the agent to remember everything it needs, but you need to actively manage the mechanics of that memory. This means understanding the cache lifetime and ensuring continuous interaction or explicit session management.

For designers building agentic workflows, optimizing for cache reads means designing UIs that encourage continuous engagement or provide clear indicators for session state. Avoid long pauses without explicit user action or system prompts to keep the cache alive.

The cheapest token that costs the most.