Same model, 68% to 88% fewer tokens: Antigravity CLI vs Claude Code vs Codex, measured

Update (Oct 23): I added more game runs, so every tool now has 5 (52 runs in total). The same Opus 5.5 now reads 71% to 89% fewer input tokens in Antigravity than in Claude Code (was 68% to 88%), and