Know what your cache is doing.

Measure cache reads, check what changes with your prefix, and compare the cost of repeated Claude prompts. Try the sample without a token, or request access for live runs.

MetaBench helps developers measure prompt-cache behavior and costs in Claude applications. The current version introduces Cache Lab, with controlled request replay, token-usage reports and JSON/CSV exports.

The developer preview implements a baseline, a cache-enabled first pass, an identical replay and an optional changed-prefix request. Reports expose token counters, timings and estimated costs.

Live Claude Sonnet 4.6 integration has been tested on Cloudflare Workers: 3,579 cached tokens read on replay. The sample uses prepared results. Email us for a token to run your own prompts.

Built on the MIT-licensed MetaBench codebase by nMaroulis. The leaderboard is not available in this version.