Most teams evaluate AI coding agents by their monthly bill and gut feeling, but neither reveals what the agent actually did. An experiment using Claude Code across three coding approaches (open/default, TDD, and skip-permissions) with OpenTelemetry telemetry piped to an observability backend reveals that total cost is a misleading metric. Per-turn cost and cache hit rate are far more informative signals. Cache invalidation events — where the agent rewrites its own plan mid-session — are the primary cost driver, not code generation itself. About 63% of spend is context re-read, ~25% is cache writes, and only ~11% is generated output. The key cost levers are: keeping context lean, scoping sessions tightly, matching model size to task complexity, and generating in proportion to need. These findings were validated across 20,000+ real Claude Code sessions over 90 days.
