PlanBench
View original ↗At a glance
- Tests
- Planning (Executive functions): ordering steps so they reach a goal without dead ends.
- How
- Blocks start stacked in towers and must be rearranged into a goal arrangement, moving one top block at a time.
- Best AI
- 97.8%
- Humans
- —
- Our items
- 3
- Executive functions gap, overall
- +17.5
Example item
A specimen written for this page. Live items are similar but never shown, so posted answers can't spoil them.
Move one top block at a time until the towers match the goal.
(a) Start
(b) Goal (3 moves at least)
Human vs AI
What the benchmark reports, and where the faculty stands overall
Best published AI
97.8%
o1-preview on block-stacking plans (52.8% with disguised names)
Humans
—
This benchmark doesn't report a human figure.
Executive functions overall · faculty reference, 0–100
Measured on other tests, not on PlanBench.
Human: GAIA, university-educated testers. AI: GAIA, Claude Sonnet 4.5 (HAL).
How we mirror it
The same puzzle, played by tapping: pick up the top block of a tower and drop it on another. Each puzzle has a known minimum number of moves, and the obvious first move is usually a trap.
Every move is stored. Solving in the minimum, or one more, passes; extra moves show how far ahead someone planned.
Items are original, versioned, and assigned at random, so a posted answer spoils only one variant.
What we record