Arm 1

Arm 1: Claude decides every 50 ms step itself (closed Oct 25, 2026). Opus 5.5 at max effort, one fresh session per ride.

• Frozen time (the sim waits for Claude between steps): it rode every bolted setup, e-bike and pedaling, straight and slalom. Each ride worked out the textbook bike controller from scratch: it probed the bars, steered into the lean, and leaned toward the next gate. A fixed two-line steering rule rides the straights but falls on every slalom, so Claude was doing more than the textbook rule.
• Free robot, even with frozen time: it fell on both the e-bike and the pedaling slalom (one ride each). One rough arm or leg move started a chain of lost grips.
• Under time pressure it falls. In real time it fell at 2.5 s, before its first move. With a 3 s deadline per step, 2 of 8 rides finished, both motorized straights that mostly balance themselves. With untimed planning first and the deadline after, it still fell, because no reply came in under 3 s.
• It took a median 4-8 s of thinking per 50 ms step, about 100-150x slower than real time.
• In all: 30 of 39 rides reached the goal, about $489.

Verdict: given unlimited thinking time, reasoning alone rides the bolted robot, so Weng's sentence doesn't hold there. It fails on the free robot and under any deadline.

Caveat: 1-4 rides per setup.