GitHub's usage report now splits cache-read from cache-write tokens per model.

Available to Copilot Business/Enterprise admins and individual users from the AI usage page (GitHub Changelog). This is the specific number you need to know whether prompt caching is paying for itself. Aggregate token counts hide a cache that's written far more than it's hit, which is the default state of a badly-partitioned prompt. Download it before you tune your context strategy, not after.