Anthropic adds prompt caching to Claude, slashing costs up to 90%

If you build with frontier LLMs, you know the pain of re-sending an entire codebase or API spec on every single turn. You are paying full price over and over just to remind the model who it is.

Anthropic rolled out Prompt Caching on its API, cutting input costs by up to 90% and slashing latency by up to 85% for long prompts. The feature caches repetitive context prefixes across calls, so Claude does not have to re-read your documents from scratch every turn.

Sub-3-second responses on full-book prompts

The speedups on large prompts are hard to ignore. In a test querying a 100,000-token book, the time to first token plummeted from 11.5 seconds down to 2.4 seconds, an 79% latency drop, while saving 90% on input costs.

For a 10-turn conversation with an extensive system prompt, latency dropped from roughly 10 seconds to 2.5 seconds, cutting overall cost by 53%. Notion has already baked the feature into Notion AI to make responses faster and rein in operational costs.

Under the hood, developers pass a cache_control parameter in their API requests. You can define up to four breakpoints or let Claude automatically cache previous dialogue turns at 1,024-token boundaries.

Writing to the cache costs 25% more than the baseline input token rate. For Claude 3.5 Sonnet, that means $3.75 per million tokens upfront instead of the standard $3. But every cache hit after that is discounted by 90%, dropping to just $0.30 per million tokens.

The default cache lives for 5 minutes, resetting each time it gets hit, alongside a 1-hour retention tier. One catch: TTL starts when the request begins, meaning a long streaming generation eats directly into your reuse window.

Why it matters

We think prompt caching is easily one of the most consequential API optimizations this year, because it completely changes the unit economics of AI agents.

Up until now, building multi-step agents or passing dozens of few-shot examples was a luxury reserved for well-funded startups. Every tool call and reasoning step bloated the context window and multiplied the bill. By dropping subsequent read prices to $0.30 per million tokens, Anthropic makes heavy context practical for everyday indie hackers and production apps alike.

For the broader industry, this shifts the battleground from raw context window size to real-world context efficiency. Having a 1-million-token window is useless if each query costs several dollars and takes 15 seconds to reply. Anthropic just set the benchmark for operational viability, putting pressure on competitors to optimize inference architectures rather than just bragging about parameters.

The catch lies in your traffic patterns. If your app sees sporadic queries spaced more than five minutes apart, you risk paying the 25% cache-write premium repeatedly without reaping the 90% discount. To make it pay off, developers will need to design their agent loops and user sessions very deliberately around cache lifetimes.