Antigravity CLI vs Claude Code vs Codex: startup tokens, MCP tools and the same coding task

Measured 2026-10-22 on Windows 11 with Claude Code 2.1.295 (Opus 5.5), Codex CLI 0.162.0-alpha.17.2 (gpt-6-luna) and Google's Antigravity CLI 1.4.5 (agy) with Gemini 4 Argon and with Claude Opus 5.5. 44 runs. License: CC BY 4.0. Page with tables, chart and method: https://tokensave.app/data/antigravity-cli-vs-claude-code-vs-codex Key numbers Startup ("Reply with the single word OK.", no MCP): Claude Code 35,781 input tokens, Codex 14,818, Antigravity + Gemini 4 Argon 6,745, Antigravity + Opus 5.5 5,326 (+5,167 cache reads reported separately). Same brick-breaker game task, average of 3 runs: Claude Code + Opus 5.5 78,671 input / 7,313 output / 58 s; Codex 55,680 / 4,563 / 68 s; Antigravity + Gemini 13,041 (+6,796 cache) / 6,373 / 81 s; Antigravity + Opus 5.5 9,822 (+15,381 cache) / 4,378 / 42 s. Same model, Opus 5.5: 68% to 88% fewer input tokens in Antigravity than in Claude Code (high end counts only reported input, low end adds the separately reported cache reads). 30 dummy MCP tools added +388 tokens (Claude Code), +0 (Codex), +582 (Antigravity + Gemini), +417 (Antigravity + Opus). All 12 games worked in a headless-browser check (score, level 2, lives 3-2-1-0). The automatic check could not read the score in two Codex games; manual checks confirmed scoring (see check_note). How to read Antigravity's numbers agy --output-format json reports input_tokens, output_tokens, thinking_tokens, cache_read_tokens and total_tokens (= input + output). Whether cache reads are part of input is undocumented; in the Opus runs cache reads exceed input, so they are not contained there. Treat input + cache_read as an upper bound. Files antigravity-cli-vs-claude-code-vs-codex.csv: one row per run (tool, test, MCP tools, run, reported input, cache reads, output, seconds, API cost for Claude Code, file lines, game checks). antigravity-cli-vs-claude-code-vs-codex.json: the same with setup and notes. antigravity-cli-vs-claude-code-vs-codex-chart.png Raw JSON output of every run, every generated game, the runner, the MCP test server and the game checkers: https://github.com/johnisking/token-cost-by-language/tree/main/antigravity-cli (Windows user name masked as USER). Earlier Claude Code vs Codex dataset: https://tokensave.app/data/claude-code-vs-codex-tokens