Haejo
Benchmarks
How many tokens and how much time an agent spends on the same job, what replaying a saved Flow costs, and which web actions run in the background in Safari. Each measurement states its conditions once.
53.5% fewer input tokens on the same tasks
Codex’s built-in computer use and Haejo were given the same three tasks, each run once in a fresh conversation. The numbers cover everything from working out the steps to running them and the final reply.
| Measure | Haejo | Codex built-in computer use |
|---|---|---|
| Tasks completed | 3/3 | 3/3 |
| Total input tokens | 201,44753.5% fewer | 433,312 |
| Input tokens excluding cache reads | 22,11956.3% fewer | 50,592 |
| Output tokens | 1,605 | 1,604 |
| Mean time | 42.10 s8.8% shorter | 46.17 s |
2026-10-14 · a development build of Haejo · Codex CLI 0.161.0, requested model gpt-6.1-sol, medium reasoning effort · tasks on a macOS test app we built: turn a bonus on, turn it off, and cope with a renamed button. On the bonus-off task Codex’s built-in computer use was faster (37.91 s against 48.55 s). The measurement record calls the other side Sky.
Replaying a saved Flow: zero model calls
An 11-step Flow (five inputs, a conditional checkbox, compute, review, completion check) was replayed with different values, and we deliberately changed the screen or made completion impossible to confirm.
| Situation | Result |
|---|---|
| 10 replays with different values | 10/10 completed · 0 model calls |
| All normal jobs | 13/13 completed |
| Two buttons changed within one job | Those steps repaired, job completed · 2 model calls |
| Completion could not be confirmed | 5/5 stopped on the spot · 0 inputs sent again |
2026-10-14 · a development build of Haejo · a macOS test app we built.
20 of 22 Safari page actions run in the background
Actions were sent to a page in the person’s own Safari and judged by the events and values the page actually received. Haejo starts with the quietest route and names the route in the result.
| Route | Actions | Your screen |
|---|---|---|
| Accessibility | 18 | Safari stays behind; pointer and keyboard untouched |
| Keys sent to that window | 2 | Safari stays behind |
| Safari in front, briefly | 2 | In front for that input only, then your app is in front again |
2026-11-12 · Safari 27.2 · macOS 27.2 · a measurement page we built. Covers links and buttons, text fields, menus, dates, scrolling, new tabs and windows, and iframes. The two that come to the front are a menu that opens on pointer-down and a menu revealed by hover.