Prompt caching allows engineering teams to store frequently used context—such as large document libraries or codebase summaries—and pay a reduced rate for subsequent requests. Across the Claude model family, costs are split into higher 'Write' fees to store the data and significantly lower 'Read' fees to retrieve it.
Haiku 4.5 provides the most aggressive savings with a Read price of $0.10 per 1M tokens. For advanced reasoning, Sonnet 5 and Opus 5 offer higher performance tiers at $0.20 and $0.50 per 1M read tokens, respectively. These rates provide a path to scale long-context applications without linear cost increases.
How do write and read fees impact margins?
Caching introduces a two-tier cost structure. You pay a premium to commit data to the cache, such as $12.50 per 1M tokens for Fable 5. Once cached, every subsequent request that references that data is charged the 'Read' rate, which for Fable 5 is $1.00 per 1M tokens. This makes caching highly effective for applications with high repeat usage of the same background data.
Which model is most efficient for high-volume caching?
Haiku 4.5 is the most cost-effective option for developers, with a Write price of $1.25 and a Read price of $0.10 per 1M tokens. If your workflow requires the higher reasoning capabilities of Opus 5, the Read cost increases to $0.50 per 1M tokens. Sonnet 5 sits in the middle, offering a balance of performance with a Read price of $0.20 per 1M tokens.
To determine if caching is delivering a true return, you must track the cache hit rate against the total time saved by end-users. NetLift helps you map these infrastructure savings back to business value, ensuring that lower API unit costs translate into a measurable net return on your AI adoption rather than just a technical optimization.
Sources
All pricing on this page comes from official vendor pages: