The best AI passes 37% of real engineering tasks. Most people would pass none

Xplore released RWE-bench today, 20 mechanical, electrical and aerospace tasks written by engineers who do this for a living. One is a diesel crankshaft that has to survive 2,400 rpm, one is routing a DDR3 memory bus on a real FPGA board, and one is CFD on a turbine vane. No model grades the results. A program checks every design against every requirement and runs physics solvers where it has to. GPT-6 Astra leads at 37%, Opus 5.5 is at about 29% and GPT-6.1 Sol at about 25%, for $10 a run. Astra and Opus cost about six times that per run for 12 and 4 more points. 37% sounds low until you ask how a person would do. Give someone off the street the same tasks and tools and they pass none, because they won’t get past drawing the part in CAD and running the simulation. An engineering graduate might get one or two in their own field after a month of work. A specialist would pass the ones in their own field, about a third, at roughly a week each. The model does some of all three fields in a few hours a run. The human time and what they could do are my own rough estimate, Xplore doesn’t publish a human score. (but i think they should next time!) Code got good fast because you can check th