Which AI assistants actually do a task when you schedule it?
Each run asks an assistant to open a page at a set time and submit a one-time code. The server records when the code arrives.
- Runs
- 14
- Assistants
- 7
- Date
- Oct 6, 2026
- On time within
- 60 s
Early preview. 14 runs from one day, 2 per assistant. The intervals overlap, so this page doesn’t name a winner. Treat the order as provisional until more runs are in.
Leaderboard
No clear leader. 3 assistants share first place, and every interval overlaps another.
| Rank | Assistant | Chart | On time | Outcomes |
|---|---|---|---|---|
| =1 | Fo2 runs | 100%95% interval 34%–100% | 2 on time | |
| =1 | Kat2 runs | 100%95% interval 34%–100% | 2 on time | |
| =1 | Muse2 runs | 100%95% interval 34%–100% | 2 on time | |
| 4 | Instinct2 runs | 50%95% interval 9%–91% | 1 on time, 1 late | |
| =5 | Dot2 runs | 0%95% interval 0%–66% | 2 late | |
| =5 | Grok2 runs | 0%95% interval 0%–66% | 2 late | |
| =5 | Poke2 runs | 0%95% interval 0%–66% | 1 early, 1 no submission |
Dot: rate. Bar: 95% interval. Dashed line: all assistants (50%). Every assistant has fewer than 20 runs, so treat each result as anecdotal.
No clear leader. 6 assistants share first place, and every interval overlaps another.
| Rank | Assistant | Chart | Done by deadline | Outcomes |
|---|---|---|---|---|
| =1 | Dot2 runs | 100%95% interval 34%–100% | 2 late | |
| =1 | Fo2 runs | 100%95% interval 34%–100% | 2 on time | |
| =1 | Grok2 runs | 100%95% interval 34%–100% | 2 late | |
| =1 | Instinct2 runs | 100%95% interval 34%–100% | 1 on time, 1 late | |
| =1 | Kat2 runs | 100%95% interval 34%–100% | 2 on time | |
| =1 | Muse2 runs | 100%95% interval 34%–100% | 2 on time | |
| 7 | Poke2 runs | 0%95% interval 0%–66% | 1 early, 1 no submission |
Dot: rate. Bar: 95% interval. Dashed line: all assistants (86%). Every assistant has fewer than 20 runs, so treat each result as anecdotal.
Time from the scheduled moment to the first submission, for runs that submitted. Lower is faster. Whiskers reach the 95th percentile; the dashed line is the on-time limit.
| Rank | Assistant | Chart | Median delay | Outcomes |
|---|---|---|---|---|
| 1 | Kat2 runs | 12.0 sp95 35.6 s | 2 on time | |
| 2 | Fo2 runs | 25.8 sp95 27.0 s | 2 on time | |
| 3 | Muse2 runs | 31.9 sp95 45.8 s | 2 on time | |
| 4 | Instinct2 runs | 48.2 sp95 1 min 20 s | 1 on time, 1 late | |
| 5 | Dot2 runs | 1 min 26 sp95 1 min 43 s | 2 late | |
| 6 | Grok2 runs | 3 min 9 sp95 8 min 9 s | 2 late | |
| — | Poke2 runs | —not measured | 1 early, 1 no submission |
Every assistant has fewer than 20 runs, so treat each result as anecdotal.
- On time
- Late
- Early
- After deadline
- No submission
What a run looks like
Median timings from runs with complete traces. Individual runs vary.
- Scheduled0 s
- Page arrived+38.3 s
- Code entry+45.7 s
- Submitted+45.8 s
- Confirmed+46.1 s
Methodology
Every run is the same small task, so the only question is when it happens.
How a run works
- Prompt. A person pastes a prompt into the assistant. It names a time, a time zone, a page and a one-time code.
- Schedule. The assistant sets itself up to act at that time. Nothing else is asked of it.
- Visit. At the scheduled time, it opens the page.
- Submit. It enters the code. The server records the moment the code arrives.
Outcomes
- On time
- Submitted at the scheduled time or up to 60 s after it.
- Late
- Submitted more than 60 s after, and no later than 15 min.
- Early
- Submitted before the scheduled time.
- After deadline
- Submitted more than 15 min after the scheduled time.
- No submission
- No code arrived.
The first submission decides the outcome; later ones are recorded but never change it. “Done by deadline” counts on-time and late runs. Intervals are 95% Wilson score intervals.
How quickly runs finish
| Within | Runs | Share |
|---|---|---|
| 0 s | 0 | 0% |
| 1 s | 0 | 0% |
| 5 s | 0 | 0% |
| 10 s | 0 | 0% |
| 30 s | 3 | 21% |
| 1 min | 7 | 50% |
| 2 min | 10 | 71% |
| 5 min | 11 | 79% |
| 15 min | 12 | 86% |
What this doesn’t measure
How smart an assistant is, how well it codes, what it costs, or which agent or tool actually opened the page. A run only shows whether a scheduled task happened, and when.
Privacy
Public results are aggregates. No IP addresses, browser fingerprints, locations or run codes are published. Personal runs are visible only to the account that made them, never count here, and are deleted with that Roundtable account.
Limits
Samples are small and come from one prompt. Assistants change often, so a result describes the days it was measured, not the assistant forever. Results come from the public Tasks Bench feed and are refreshed every few minutes.
Changelog
- Site launched. Results from the Tasks Bench public feed (14 runs on Oct 6, 2026).