We test AI coding agents so you don't have to. Claude Code, Codex CLI and Antigravity CLI each get the same complete product brief, one autonomous run and zero human help. Then we publish the app they shipped, the wall-clock time, the tokens, the cost and the bugs. 1 Prompt. 1 App. 0 Help.
A single-file HTML5 Canvas Tower Defense game requiring 60fps loop, 3 distinct tower types, upgrade/sell mechanics, wave management, particle explosions, and Web Audio API procedural sound synthesis.
Output tokens and cost come from each CLI's own run log. Codex CLI and Antigravity CLI report tokens only, so their cost is unknown. The Claude Code tower defense log kept the final message but not the usage block, so its tokens and cost are unknown.
A single-file 3D arcade kart racing game with procedural Three.js low-poly models, closed-loop stunt circuit, drift physics, tire smoke particles, 2 AI rival karts, 3-lap race system, boost pads, and real-time Web Audio sound synthesis.
Disclosure: the Claude Code pane in this recording used a copy with a 1-line fix applied by us. The unedited agent output freezes, which is why that run is now marked no-ship.
Output tokens and cost come from each CLI's own run log. Claude Code reports a list-price dollar cost; Codex CLI and Antigravity CLI report tokens only (subscription billing), so their cost is shown as unknown rather than estimated.
A single-file HTML expense tracker with an add form, a newest-first table, a category filter, localStorage persistence and a running total of all expenses. Run headlessly on Claude Code (claude-opus-5), Codex CLI (gpt-5.6-sol) and Antigravity CLI (gemini-3.1-pro-high). Two of three agents turned the grand total into a filtered subtotal.
Each agent receives an unedited, exhaustive product specification once. Zero human semantic follow-ups, zero steering, and zero debugging hints.
The agent plans, codes, inspects, and verifies on its own inside an isolated sandbox. Harness versions, models, effort settings and environments are recorded when the CLI exposes them.
We capture exact wall-clock durations, token consumption, session transcripts, and unedited code outputs so claims can be audited independently. Anything a CLI did not report is published as unknown, never estimated.