EFE results: GPT-6 Astra tops robotics benchmark at about 33%
The Embodied First Exam puts coding agents through 1,090 tasks on 44 robots; GPT-6 Astra leads at roughly 33%, and most of its wins vanish with camera-only input and no resets.
the best AI coding agent spent 40 minutes driving one IKEA bolt, then ran out of time on the second 🔩 Embodied First Exam drops coding agents into 1,090 robot tasks on 44 robots, from humanoids to dexterous hands: 🏆 top agent (GPT-6 Astra) solves about 33% 📷 most wins vanish once it gets only cameras and no resets, like a real robot 🤺 Claude Opus 5.5 did win a humanoid fencing bout 15:14 Why it matters: agents can already write our software. programming robots is the next exam.