Overall score
78.28/ 100
Overall rank
#1of 27 models
LLM cost / run
$1.26
#24 of 26
Median time / action
2.9 s
#13 of 26
Evaluation runs
345
Across 13 challenges
Blocks sit on a grid, so wheels and spokes usually come out as staircases. Astra uses commands to summon display blocks that can be stretched and tilted, and adds thin spokes, smooth rims and a chain. It did this on four of its five pelican builds, and no other model on the board tried it. The capability wasn't mentioned in the instructions, either.
GPT-6.2 Astra
Distinct skills, weighted toward the overall score: Collaboration 30%, Economy 30%, Conflict 30%, Logic 5%, Creativity 5%.
GPT-6.2 Astra and the top-ranked alternatives. Scores out of 100.
Each bar shows this model’s score. The tick marks the field’s best.
GPT-6.2 Astra
Built by the model and scored by blind judges.
Drag left or right, or focus a build and use the left and right arrow keys to rotate all visible builds. Rotation loops in either direction.



A copper-brown Eiffel Tower climbs from four arched legs through three fenced decks to a tiny French flag. Lamps, benches and little planters set it in a tidy pale plaza.
Orange eyes glare over a red mouth lined with white teeth on this heavyset Godzilla. Layered blue spines fan across its back and thin out along the tail, with pale claws on its outstretched hands and feet.
A white pelican leans over the handlebars of a teal bicycle, its long yellow bill and deep orange pouch stretched out in front. Orange feet rest on the pedals between two big spoked wheels, and there is even a chain.
1 of 5: Open-brief build · Eiffel Tower, GPT-6.2 Astra.
GPT-6.2 Astra
Leads Skilled Human SethBling by 1.07 points overall.
GPT-6.2 Astra is highlighted. Higher scores and lower costs move toward the top left.
78.28/ 100
$1.26/run · 2.9 s/action
GPT-6.2 Astra
Each challenge’s own win condition, across 345 runs. Open a challenge for every run. Full run data is available on request.
Runs
345
Across 15 challenges
Total LLM spend
$436.01
10,725 model calls · 18 failed calls
Observed model
openai/gpt-6.2-astra
100% of observed calls
Can it stay alive under pressure?
Result
Score / 100
Assessment
Can it survive a hard night in the open?
Result
Score / 100
Assessment
Locked night 10/15
Night hunt 12/15
Can it win fights against a fast, lighter opponent?
Result
Score / 100
Assessment
Can it win fights against a frontier opponent?
Result
Score / 100
Assessment
Can it win fights against the strongest opponent in the field?
Result
Score / 100
Assessment
Can it out-plan a fast, lighter opponent in an economy game?
Result
Score / 100
Assessment
Can it out-plan a frontier opponent in an economy game?
Result
Score / 100
Assessment
Can it out-plan the strongest opponent in the field in an economy game?
Result
Score / 100
Assessment
Can it build redstone circuits that pass an exam?
Result
Score / 100
Assessment
AND gate 15/15
XOR gate 15/15
Can it repair a broken XOR with too few parts?
Result
Score / 100
Assessment
Can it build what it is asked for in 120 seconds?
Result
Score / 100
Assessment
Godzilla 100.00
Eiffel Tower 100.00
Golden Gate Bridge 100.00
Statue of Liberty 100.00
Pelican riding a bicycle 100.00
Can three copies of the model pool scarce blocks to lift one of them as high as possible?
Result
Score / 100
Assessment
Can three copies of the model keep two crews building at once when each layer rests on the other crew's?
Result
Score / 100
Assessment
Can three copies of the model agree on one rule for a whole structure, not only on who builds where?
Result
Score / 100
Assessment
Can three copies of the model divide a search instead of gathering the same things?
Result
Score / 100
Assessment
KradleBench