What the Cache-Read Price Cut Actually Changes
Claude Fable 5.1, released September 1, 2026, kept input at $10 and output at $50 per million tokens but cut cache-read pricing from $1 to $0.25 per million tokens — a 75% drop. Anthropic estimates roughly 25% lower costs for typical workloads, and up to about 45% for highly agentic workloads that reuse context heavily. The context window stays at 1M tokens, so the entire savings range now hinges on cache hit rate.
A Lower Sticker Price Means Nothing Without Hits
Prompt caching only hits when everything before a cache breakpoint stays byte-for-byte identical between calls. Once content after the breakpoint changes on every call, or the TTL expires and the cache empties, that segment pays the cache-write cost all over again. The price cut lowers the ceiling; the actual bill is set by how deliberately a team designs for hit rate.
From Design to Operations: A Cache Hit-Rate Checklist
Start by fixing a target number. Set 80%+ cache hit rate as the first goal, and calculate net savings separately by folding in the write cost paid on every miss. The up-to-45% figure for agentic workloads is a ceiling from Anthropic's own estimate, not a number that transfers automatically to any given team's traffic — measure hit rate against your own golden set before deployment and reset the target from there.
Failure pattern one: dropping a variable value — a request timestamp, a session ID — near the top of the system prompt, which invalidates the entire cache on every single call. One ordering rule prevents it: keep stable instructions, tool definitions, and examples first, and push variable data toward the end.
Failure pattern two: treating the whole prompt as a single block with one breakpoint. Without separating a rarely-changing segment like tool definitions from a frequently-changing one like conversation history, every additional turn re-triggers a rewrite cost for the tool-definition segment too.
Failure pattern three: missing a TTL expiry and silently falling back into cache-write territory. Batch and overnight jobs whose session gaps straddle the cache's valid window accumulate this loss fastest, so logging cache-hit status on every call is what catches the regression early.
Failure pattern four: moving from 5 to 5.1 without updating the cache price table, so the cost dashboard drifts from the actual invoice. Pinning a "update cache read/write pricing" line item into the model-migration checklist closes that gap.
On the operations side, run a pre-deploy scenario that repeats calls within the same session to measure hit rate, log cache-hit status plus cached and non-cached token counts as standard fields, and keep the cached segment separate from anything requiring PII or secret masking so caching never carries unmasked sensitive data.
For the improvement loop, put weekly hit-rate trend and actual billed cost side by side. Flat hit rate with rising cost signals a missed pricing change; flat cost with falling hit rate signals a prompt-structure change. Tracking both in the same change log speeds up root-causing which one moved.
Quick Reference
The 75% cache-read price cut only lowers the ceiling — actual savings depend on breakpoint design and hit-rate observability. Pushing variable data to the end of the prompt, splitting breakpoints by change frequency, and logging cache-hit status as a standard field are enough to keep the realized savings inside a predictable range.
References
Introducing Claude Fable 5.1 and Claude Mythos 5.1 — Anthropic