Dream Labs has opened Dream API, a public research beta that runs Dream-1, a World Action Model for two-arm robots with parallel grippers. Developers and robot makers send camera views plus a plain-language instruction and get motor commands back, including on robots the model was not trained on, without demonstrations from that body and without training a policy themselves. Hosted inference costs $3 an hour for Dream-1 and $2 an hour for the lighter Dream-1-flash. Capable two-arm setups now sell for a few thousand dollars, but getting a new robot to do useful work has still meant data from that exact machine and months of specialist effort. Dream-1 is trained on video of people doing everyday tasks, predicts how a scene will change, and chooses actions to match that prediction. The team says that is what lets one model attempt tasks and places it was never shown. Dream Labs reports zero-shot tests on 27 embodiments it had never trained on in simulation, including a PR2 from 2008, and real rollouts on partner arms including MakerMods' Maker Arm and Metal Arm and Anvil Robotics' OpenYAM. Robocurve's independent eval on RoboDojo-RC Tier 1 scored a mean of 59 per task attempt, against 48 for the next best models, and Dream Labs says that report also shows 16 times lower cost per task attempt than Gemini 4 Argon. Joel Jang, co-founder and CEO, told Forbes that zero-shot pick-and-place succeeds about 80% of the time, with other tasks in the 50% to 60% range, and that the company is not aiming at factory lines that need near-perfect reliability. For now the model covers tabletop work such as picking, placing, and folding laundry, with wheeled robots and dexterous hands planned later. People without hardware can prompt real robots in the San Francisco office through the browser-based Robot Playground. The company was founded in 2026 in San Francisco by former NVIDIA GR00T researchers Joel Jang, Scott Reed, Youliang Tan, and Yunhao Ge, the group that introduced World Action Models with the DreamZero paper in February. Alongside the API, Dream Labs announced a $90 million seed round led by B Capital and Kindred Ventures. It says it works with more than 20 hardware partners, on robots roughly in the $2,000 to $10,000 range, and that a few hundred demonstrations can adapt the model when a specific body needs its own data.