Auto-caching saved $131,743.53 over 30 days. Support triage reused its context across 1.63M requests. Savings are net of cache-write premiums, compared with the same traffic without Caveman caching.
Routing saved $4,805.19 over 30 days. IT service desk handled 186,420 requests for $1,570.37, versus $6,375.56 on Sonnet 4.6 alone. Routine requests went to Flash 2.5; 25,539 complex requests stayed on Sonnet 4.6.
Tool bloat removed: 89.4M tokens / day. Procurement policy checks carried 38,672 tool-context tokens per run. Caveman reduced that to 7,248 across 2,846 daily runs, saving $268.30 / day at an assumed $3 per million uncached input tokens.
Quality retained in the sample: 1,207 / 1,248 passing cases before and after on the procurement regression suite.
Illustrative results, not executed benchmarks. Routing and tool-bloat scenarios are independent of the cache-only totals below.