Anthropic launches prompt caching in the Claude API

Anthropic added prompt caching to the Claude API, letting developers store large, frequently reused blocks of context — a long document, a detailed system prompt, a set of few-shot examples — so later calls referencing the same material are not re-processed from scratch. The feature launched in public beta for Claude 3.5 Sonnet and Claude 3 Haiku, with Claude 3 Opus support following shortly after, all within the 200,000-token context window Anthropic had introduced earlier that year.

Anthropic reported cost reductions of up to 90% and latency reductions of up to 85% for prompts drawing on a large cached block, citing a worked example of a 100,000-token cached prompt cutting response time by 79%. The pricing reflected the trade-off directly: writing new content to the cache cost 25% more than a standard input token, but reading from it cost only 10% of the standard rate, so the saving accrued only on repeated use of the same context.

The company positioned the feature for conversational agents that carry long-running context, coding assistants working across a large codebase, and agentic search tasks that repeatedly re-query the same background material — use cases where the same tokens would otherwise be billed and reprocessed on every call. Competitors offered comparable mechanisms within the following months, making cached-context pricing a standard feature of frontier model APIs rather than a distinctive one.