Caveman vs Ponytail vs RTK: Claude Code token comparison

The speaker compares Caveman, Ponytail and RTK with plain Claude Code by implementing the same transcription vocabulary feature five times. Each setup starts from one frozen commit and uses Opus 5 at medium effort in plan mode. Token totals include the main session and all sub-agents. Each setup gets only one run, so the results describe this experiment rather than a general ranking of coding assistant tools.

Ponytail in ultra mode used 18.1 million tokens, about 39% of the 46.3 million baseline. Caveman used 19.1 million, RTK used 27.6 million, and the combined setup used 27.7 million. Combining all three finished fastest at 15 minutes and 48 seconds, but did not improve token savings over Ponytail alone.

The tools act differently. Caveman injects prompt style instructions; this test excludes its gateway and separate input compression engine. Ponytail adds minimalist coding rules to the main session and sub-agents. RTK replaces supported bash commands with equivalents that produce less output.

The speaker reports that every implementation passed its tests and built. Later reviews gave Caveman and the combined setup 9/10; the others scored 8/10. Plain Claude also left a visible Return-key bug. Differences in effort levels, token accounting and tool versions limit comparisons with the cited JetBrains benchmarks. The speaker keeps Ponytail and RTK enabled.