What Actually Burns Your Claude Code Tokens — Measured, Not Guessed

I ran the same bug-fix task through Claude Code four times — same model, same fixture — changing only how much context I gave it upfront: a full repo dump, targeted file references, a minimal one-line instruction, and iterative back-and-forth clarification. All four succeeded (3/3 hidden tests, 3/3 visible tests each). Token usage is read straight from the tool's own real usage output, nothing estimated. Dollar costs are ESTIMATED from current published Anthropic pricing (verified at production time) and labeled as estimates every time they appear on screen. The shortest prompt was not the cheapest. Full context dump: 134,764 billable tokens, $0.171. Targeted references: 221,414 tokens, $0.209 — more expensive than dumping the whole repo, because a "helpfully scoped" prompt still needed Claude to explore and re-read files, and that exploration cost more in cache churn than just handing over everything up front. Minimal one-line instruction: about the same as targeted, $0.205. Iterative 3-turn back-and-forth: the most expensive by a wide margin, $0.418 — roughly 2.4x the cheapest option, for the identical outcome. One more real number this video doesn't hide: Claude Code's own built-in cost display was pricing this session's Sonnet 5 usage at the previous generation's (Sonnet 4.6) rate card — a uniform ~1.5x overstatement versus the currently published price. Shown on screen, not corrected quietly. Every command and usage figure shown is real, captured through this channel's evidence-ledger tooling. Full harness, logs, and pricing source: https://same-task-bench.pages.dev Narration is AI-generated (local Piper TTS, en_US-norman-medium, public domain). Chapters: 0:00 Cold open — same task, same model, cost differed 1:21 Rules of the run — 4 strategies, 1 model 2:01 Strategy A — full repo dump 2:44 Strategy B — targeted file references 3:20 Strategy C — iterative, 3 turns 3:56 Strategy D — minimal one-line instruction 4:24 All 4 strategies pass — cost roll-up 4:46 Where the money actually goes 6:08 The pricing discrepancy — tool vs. live rate card 7:28 What actually drove the cost 7:52 Scope limit, stated on screen 8:22 Channel promise and endcard