AI startup Decart unveiled Oasis 3 on June 10, 2026 — a real-time world model that generates photorealistic driving environments and is now available to developers via API at $0.02 per second, with enterprise pricing available depending on use case.
The model is initially aimed at autonomous vehicle companies that need to simulate rare driving scenarios at scale. Decart plans to expand into robotics and other physical AI applications. Oasis 3 generates physically accurate, multi-camera environments — one front-facing and two side-facing — and allows developers to generate scenarios indefinitely, rather than offering capped demos or research previews.
Decart co-founder and CEO Dean Leitersdorf said the launch is designed to foster a developer ecosystem around world models. “It’s going to be the first usable world model that people can actually program on top of,” he said. The company already has more than 100,000 developers building on its real-time video model, Lucy, primarily in e-commerce and live streaming. Oasis 3 is built on that same foundation model.
The release comes weeks after Decart raised $300 million, lifting its valuation to nearly $4 billion. Strategic investors in the round included Toyota, Adobe, eBay, and existing investor Nvidia. Leitersdorf attributed the raise to strong demand across e-commerce, live streaming, and physical AI.
Oasis 3’s efficiency stems from Decart’s DOS (Decart Optimization Stack), software that optimizes model performance across Nvidia, Amazon, and Google hardware. Leitersdorf said this vertical integration makes the models “more than an order of magnitude cheaper” to run than competitors, and that the company has spent “drastically less” than $100 million in its lifetime.
The model does have notable limitations. Because Oasis 3 is auto-regressive — generating one frame at a time based on prior frames — scene consistency degrades over extended sessions as the context window fills. In testing, the model also allowed vehicles to pass through other objects, meaning physical interactions are not yet accurately simulated. Leitersdorf described this as “a major research problem that we’re cracking now” and said improvements to memory length are in development. A future version may allow world generation from video input rather than a text prompt.
Source: TechCrunch