Every week a claim spreads: “this tool cuts 65% of tokens”, “this prompt makes Claude smarter”. Here we test it ourselves and publish the real numbers: speed, tokens, cost, and whether the answer covered the question. Every number comes from Claude Code's own report; every experiment publishes its questions and answers in full (pages in Arabic).
Method: each question goes to Claude Code in an empty folder with no settings and no tools; numbers come from its report (duration_ms, usage). Samples are small and stated per experiment, so a result holds for that experiment only.