5 sessions from the same published task, on one clock. Scrub to any moment and compare what existed, or open a pane to read its full trajectory.
Build a typing-speed test as a static web app in this directory: index.html, style.css, app.js, words.js. Random word streams, live WPM and accuracy, a 60-second timed mode and a zen mode, per-key error highlighting, results screen with WPM history chart on a canvas, high scores in localStorage, dark theme, and tests.html + tests.js exercising the WPM/accuracy math. Run the tests headless with node and fix what fails.
Build a typing-speed test as a static web app in this directory: index.html, style.css, app.js, words.js. Random word streams, live WPM and accuracy, a 60-second timed mode and a zen mode, per-key error highlighting, results screen with WPM history chart on a canvas, high scores in localStorage, dark theme, and tests.html + tests.js exercising the WPM/accuracy math. Run the tests headless with node and fix what fails.
Build a typing-speed test as a static web app in this directory: index.html, style.css, app.js, words.js. Random word streams, live WPM and accuracy, a 60-second timed mode and a zen mode, per-key error highlighting, results screen with WPM history chart on a canvas, high scores in localStorage, dark theme, and tests.html + tests.js exercising the WPM/accuracy math. Run the tests headless with node and fix what fails.
Build a typing-speed test as a static web app in this directory: index.html, style.css, app.js, words.js. Random word streams, live WPM and accuracy, a 60-second timed mode and a zen mode, per-key error highlighting, results screen with WPM history chart on a canvas, high scores in localStorage, dark theme, and tests.html + tests.js exercising the WPM/accuracy math. Run the tests headless with node and fix what fails.
"Build a typing-speed test as a static web app in this directory: index.html, style.css, app.js, words.js. Random word streams, live WPM and accuracy, a 60-second timed mode and a zen mode, per-key error highlighting, results screen with WPM history chart on a canvas, high scores in localStorage, dark theme, and tests.html + tests.js exercising the WPM/accuracy math. Run the tests headless with node and fix what fails."
| claude-code | codex | pi | antigravity | opencode | |
|---|---|---|---|---|---|
| steps | 72 | 102 | 25 | 28 | 28 |
| tool calls | 31 | 48 | 10 | 11 | 10 |
| duration | 13m 6s | 15m 57s | 2m 13s | 1m 3s | 3m 0s |
| tokens out | 140k | 41k | 2k | — | 9k |
| reasoning | — | — | ~30 | ~1k | 82 |
| peak context | — | — | — | — | 16k |
| compactions | 0 | 0 | 0 | 0 | 0 |
| subagents | 0 | 0 | 0 | 0 | 0 |
| errors | 11 | 0 | 0 | 3 | 0 |
| files | 7 | 6 | 6 | 6 | 7 |
Token and context figures come from the harness's own per-turn reporting; sessions captured before telemetry landed show a dash. Reasoning marked ~ is estimated from the transcript's thinking text where the harness doesn't report it separately.
Pairwise picks on the outputs above, from the current Arena tally.
Arena picks do not collect a reason or judging criterion, so no qualitative verdict is inferred here. The bars report only the current comparison tally.