Median realized reuse ratio

The median share of input tokens served from cache once provider caching is configured through Zumik.

All trends

+14 % of input tokens over window

Jan

Feb

Mar

Apr

May

Jun

Measured in % of input tokens.

What this trend means.

Realized reuse keeps climbing as prompt ordering improves. The ceiling is opportunity; the gap to it is usually prompt construction, not infrastructure.

Tie-in

Trends describe the population; a diagnostic measures you. The two together tell you whether you are ahead of or behind the curve on reuse.

Other trend series.

Workloads with a >10k-token stable prefix

% of workloads

Routing mix by alias family

% of requests

Share of non-interactive tokens on batch tiers

% of background tokens

Measure your reuse against the trend.

Run a diagnostic and see where your workload sits relative to these aggregate patterns.

Run a diagnostic The benchmark