AI BriefingToken Economics: The 62.5-Minute Rule for Claude Caches
Key point
It's advantageous to keep a Claude cache alive if you plan to write to it again within 62.5 minutes.
Details
Anthropic's prompt caching charges 5-minute writes, 1-hour writes, and reads/refreshes at different multiples of the input token price. A request that reads a live cache is billed while simultaneously resetting the TTL back to 5 minutes.
- 5-minute cache write: 1.25x
- 1-hour cache write: 2x
- Read/refresh: 0.10x
On Opus 4.7, a 100K token prefix costs $0.625 for a 5-minute cache write and $0.05 for a read. The cost of keeping a cache alive for T minutes is W + R * floor(T / 5), and the cost of rewriting is 2W, which puts the break-even point at 62.5 minutes.
This threshold stays the same when switching to Sonnet 4.6 and Haiku 4.5, and it doesn't change whether the prefix is 5K tokens or 500K tokens. In practice, the boundary is stepped because of the 5-minute refresh increments, but refreshing is always advantageous within roughly 1 hour. Ultimately, if there's a chance you'll write again within 62.5 minutes, keep it alive; if longer than that, it's better to let it expire and rewrite when needed.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
US Appeals Court Upholds Pentagon Blacklisting of Anthropic as Supply Chain Risk