Prompt Caching

Prompt caching is the Claude mechanism that lets you designate big chunks of repeated text as cacheable so Anthropic stores them and charges you far less to reuse them in follow up messages. Instead of paying the full input token rate on every single turn of a long conversation you pay once to write the cache and then a discounted rate to read it back. Claude 4.7 introduced a one hour cache tier that sits above the previous five minute maximum and reduced cache read costs by thirty percent. The API exposes specific markers that tell the system where your cacheable prefix ends. That prefix can include system instructions, full codebases, detailed rubrics, brand guidelines, or anything else that stays static across many turns. The model still sees the full prompt. The only difference is how much you pay for it. This change alone moved Claude from nice to have for prototypes to the default choice for any production agent that runs longer than a few minutes.

Prompt caching is not context summarization. It does not shrink your prompt or change what the model attends to. It is not a replacement for RAG when your data changes too frequently to cache. It is not free. Even with the discount you still pay for every read and the first write is full price. It is not automatic. Your code must explicitly request cache creation and handle cases where the cache has expired or been invalidated. It is not a silver bullet for bad prompts. If your underlying instructions are sloppy the cache just lets you fail consistently at lower cost. And it is not identical across providers. What works on Claude 4.7 will not map one to one with OpenAIs cache API or whatever Gemini calls their version this week.

The real proof lives in shipping products. Cursor integrated prompt caching into their 2026 agent mode on Sonnet 4.7. Every composer session begins with a cached block containing the entire repository map, open GitHub issues, the teams design system token values, and a detailed prompt that defines their engineering culture. That block averages two hundred twenty thousand tokens. Before caching those tokens cost them twenty eight cents per interaction at scale. After the one hour tier and the thirty percent read discount the cost per interaction fell to eight cents. The team used the savings to extend agent sessions from forty five minutes to four full hours without raising prices. Granola did the same for their Haiku 4.7 transcription service. Their cached prompt contains speaker voice profiles for every customer, historical examples of perfectly structured notes, and a dense JSON output schema with thirty seven validation rules. Meetings that used to cost nine cents to structure now cost two cents. The company flipped from losing money on heavy users to healthy margins and immediately launched the unlimited plan that crushed their competitors. Linear built their prioritization engine on Opus 4.7 with a cached company strategy document that includes the last three years of OKRs, product principles, and a fifty point scoring rubric for every issue. The cache lives for the full one hour and serves hundreds of ranking requests per hour during peak planning weeks. Their monthly Claude bill dropped forty one percent even as they doubled the number of issues processed. A design team at a fintech startup built an AI design auditor that caches their complete design system documentation, WCAG accessibility guidelines tailored to their product, and before and after examples of good versus bad UI patterns. The one hundred eighty thousand token cache lets designers get real time feedback during their Figma sessions instead of waiting for overnight batch jobs. They measured a seventy two percent reduction in token spend and a thirty eight percent increase in design quality scores because the critiques became consistent instead of varying wildly when the model forgot parts of the rubric. Devin uses prompt caching across all three tiers for their ten hour autonomous coding runs. The system prompt with project goals and acceptance criteria stays cached for the entire duration. Haiku handles quick lookups against the cache. Sonnet does the bulk of the editing. Opus steps in only for hard architectural decisions. The combined approach lets them run tasks that would have cost hundreds of dollars on 4.6 for a fraction of that price on 4.7. Brainy Papers runs its rubric based eval pipeline on the same pattern. The full fifteen criterion scoring guide with concrete examples for every level of design critique gets cached for one hour batches. Ten thousand outputs get scored for the price that used to buy two thousand.

Use prompt caching any time your agent or tool repeats the same large context across many interactions inside a single session. Customer support agents that load the full knowledge base upfront, code review tools that judge every diff against the same linting bible, long running research agents that keep the same background material for hours, and design feedback loops that always start from your brand book and design principles are the sweet spot. The math works best when that repeated context exceeds one hundred thousand tokens and the session contains at least eight to ten turns. That is where the thirty percent discount compounds into savings that actually move your unit economics. Combine it with tiered routing. Cache the prompt once on the expensive Opus call then let cheaper Sonnet and Haiku models read from the same cache for the high volume work. The 4.7 family makes this trivial because every variant supports the identical caching surface.

Stay away from prompt caching when your use case involves completely unique prompts on every call. One shot image analysis, novel copywriting, or exploratory brainstorming sessions see zero benefit and just add code complexity for no return. Avoid it when the context mutates constantly like an agent that rewrites a large document and then needs the new version for the next step. Cache invalidation in those cases requires careful engineering that often costs more time than the token savings are worth. Do not default to it for simple chatbots that finish in under five turns. The overhead of managing cache keys outweighs the minimal discount you would receive. Finally skip it if you are primarily on GPT 5.5 or Gemini 3 unless you have run the exact numbers because their cache implementations have different lifetimes, pricing, and reliability profiles that can flip the math against you.

Prompt caching is the lever that turns expensive long horizon agents and consistent design critique tools from margin killers into profitable products.