Dream Labs released the Dream-1 API to control bimanual robots without target-robot demonstrations or fine-tuning. The model jointly predicts video and actions and was evaluated on 27 simulated and four real robot embodiments. It reports substantial relative improvements over comparison models on some YAM tasks, yet full-task success on that robot was 6% for tool use and 40% for pick-and-place. Relative gains must not be mistaken for absolute success rates. Unlike the earlier DreamZero, which used 30 minutes of adaptation data for a new embodiment, Dream-1 claims zero-shot transfer; inverse-kinematics quality and a safe test environment still matter.