AsapiYour AI robot operator
We built a workshop in a physics simulation: two robot arms, boards, screws and a drill, where a grip holds only by friction, so a loose one lets things slip. Ten jobs range from setting down a beam to building a dog house's frame and roof, with every board placed exactly and, in seven, screwed in. In the eleventh, the AI swings a bat at ten pitches, seeing only what a batter would see.
11 jobs · 4 AIs tested. GPT‑6.1 Sol is the only AI tested on most of them.
The eleven jobs
Every AI on every job, each a dot from 0 to 100. Open a job to see what the best AI got right and wrong.
Move a beam with two robot arms (1 AI tested)
Working two robot arms in a physics simulation, where a loose grip lets things slip, the AI has to pick up one flat beam and set it down exactly inside its outline.
What the best AI did
- Picked up the beam and set it down near its outline in under three minutes, and the first check said it had succeeded.
- Under the stricter check behind these scores, a corner stuck out 16 mm, so the move didn't count.
How each AI scored
Turn and place three boards (1 AI tested)
With two robot arms in a physics simulation, the AI has to turn three boards to new angles and set each one down exactly inside its outline.
What the best AI did
- Turned all three boards and set them down near their outlines, and the first check said it had succeeded.
- Under the stricter check, each stuck out 16 to 44 mm at a corner, so none of them counted.
How each AI scored
Place two boards (1 AI tested)
With two robot arms in a physics simulation, the AI has to pick up two boards and set each one down exactly inside its outline.
What the best AI did
- Picked up, carried and set down both boards near their outlines.
- Used up its simulation time still nudging them, with one board 4 mm and the other 9 mm outside their outlines, so neither counted.
How each AI scored
Screw two overlapping boards together (1 AI tested)
With two robot arms, the AI has to overlap two boards and fix them with three screws, picking up each screw, lining it up with a drill and driving it home.
What the best AI did
- Picked up both boards and set them down at the joint.
- Ran out of its 20 minutes without driving a screw, so the boards were never joined.
How each AI scored
Build a three‑board corner (1 AI tested)
With two robot arms, a drill and six screws, the AI has to build a corner from three boards, each set exactly in place and fixed.
What the best AI did
- Moved all three boards to the corner.
- Ran out of its 20 minutes with the boards 38 to 61 mm out of place and no screw in, so the corner was never built.
How each AI scored
Build a three‑board bridge (1 AI tested)
With two robot arms, a drill and six screws, the AI has to build a small bridge from three boards, each set exactly in place and fixed.
What the best AI did
- Laid all three boards out as a bridge and picked up the drill.
- Never got a grip on a screw and left every board outside its outline, so the bridge was never built.
How each AI scored
Build a five‑board grid (1 AI tested)
With two robot arms, a drill and 18 screws, the AI has to lay five boards into a grid, each set exactly in place and fixed.
What the best AI did
- Moved all five boards towards the grid.
- One board ended 16 cm out of place and no screw went in before its time ran out, so the grid was never built.
How each AI scored
Build a nine‑board grid (1 AI tested)
With two robot arms, a drill and 42 screws, the AI has to build a grid from nine boards, each set exactly in place and fixed.
What the best AI did
- Moved several of the shorter crosspieces towards the grid.
- Kept dropping the long rails and never got as far as the screws, so the grid was never built.
How each AI scored
Build a slatted bench (1 AI tested)
With two robot arms, a drill and 42 screws, the AI has to build a slatted bench from 11 boards.
What the best AI did
- Moved slats and legs towards the bench, and picked up the drill.
- Dropped pieces on the way, and each of the four screws it grabbed slipped out, so the bench was never built.
How each AI scored
Build a dog house's frame and roof (1 AI tested)
With two robot arms, a drill and 90 screws, the AI has to build a dog house's frame and roof from 19 boards, the biggest of the ten builds.
What the best AI did
- Took several roof slats off the pile.
- Spent much of its time trying to reach the boards underneath and never got to the tools, so no part of the dog house went up.
How each AI scored
Bat ten pitches with two robot arms (4 AIs tested)
Two robot arms hold a baseball bat, and the AI, seeing only what a batter would see, moves their eight joints to hit ten pitches as far as it can.
What the best AIs did
- GPT‑6.1 Sol hit two of the ten pitches, more than any other AI.
- Both hits went foul, outside the lines, so neither earned points for distance.
How each AI scored
Who did best
OpenAI's GPT‑6.1 Sol is the only AI we've tested on the builds. It moved boards on most, but every board ended outside its outline and no screw went in, so no build was complete. At batting, it hit two pitches and GPT‑6 Astra and GPT‑6 Luna one each, all foul; GPT‑6.1 Terra missed all ten. Opus 5.5, Fable 5.1 and Sonnet 5.5 stopped early, two of them when the account they ran on hit its usage limit, so they have no score.
What the eleven jobs showed
What they do well
Gripping and carrying boards went well.
GPT‑6.1 Sol picked up, carried and set down boards on most builds, and turned three boards to new angles before setting them down.
Where they fall short
Exact placement is out of reach so far.
Every board GPT‑6.1 Sol set down stuck out of its outline, by 4 mm at the closest, so not one build counted. Each outline is the board's exact size, and we haven't yet shown that a board can be set down that precisely.
Handling small parts is harder still.
It picked up a drill on some builds, but every screw it grabbed slipped before it could be driven, so no screw went into any build.
Reacting while the ball is moving is hard too.
At batting, the pitch keeps coming while the AI thinks, though slowed to a fiftieth of its speed. The four AIs that faced all ten pitches hit four between them, every one foul, so none earned points for distance.
How these jobs were set up and scored
- The workshop
- A physics simulation of two robot arms, boards, a drill and screws, where a grip holds only by friction. On the builds, the AI reads exact positions, moves the robot arms by typed commands, and the simulation waits while it thinks. At batting, it sees only pictures from the batter's point of view.
- Scores
- A build counts only when every corner of every board sits inside its outline and every screw is in: 100, or 0. Batting earns a point for each hit, plus a point for each metre a ball carries inside the lines before it lands; a foul, outside the lines, earns only the hit's point. We show it out of 100, with 1,000 points as 100.
- A stricter check
- The builds first used a looser check, which accepted boards a few centimetres off. The scores here use the strict one, where each outline is the board's exact size. On the first five builds, the AI saw the looser target while it worked, and on two it stopped when the old check said it had succeeded.
- Time and cost
- We tested GPT‑6.1 Sol once on each build, with 20 minutes each, and four AIs once at batting; three more started but stopped early, so they have no score. The builds took 2 to 20 minutes. They ran on subscriptions, so cost wasn't recorded.
Every AI's score on every job
| AI | Move a beam with two robot arms | Turn and place three boards | Place two boards | Screw two overlapping boards together | Build a three‑board corner | Build a three‑board bridge | Build a five‑board grid | Build a nine‑board grid | Build a slatted bench | Build a dog house's frame and roof | Bat ten pitches with two robot arms |
|---|---|---|---|---|---|---|---|---|---|---|---|
| GPT‑6.1 SolOpenAI | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| GPT‑6 AstraOpenAI | – | – | – | – | – | – | – | – | – | – | 0 |
| GPT‑6 LunaOpenAI | – | – | – | – | – | – | – | – | – | – | 0 |
| GPT‑6.1 TerraOpenAI | – | – | – | – | – | – | – | – | – | – | 0 |
These are the scores out of 100, with each AI tested once. A build counts only when it's complete, and batting shows 1,000 points as 100. A dash means we didn't test that AI on that job.