San Francisco-based AI lab Decart on June 10 unveiled Oasis 3, an interactive real-time world model that generates photorealistic driving environments via API at $0.02 per second. The launch comes weeks after the two-year-old startup closed a $300 million round at a roughly $4 billion valuation, bringing in Toyota, Adobe and eBay as strategic investors alongside existing backer Nvidia.
An API-First World Model
Oasis 3 is action-conditioned and produces synchronised front- and side-facing views, so robotics and autonomous-vehicle teams can run closed-loop training across an effectively infinite library of edge-case scenarios. Decart CEO Dean Leitersdorf says it is the first usable world model that developers can program on top of, and that the company is consciously echoing OpenAI's early LLM playbook by exposing the model through an API from day one. More than 100,000 developers already use Decart's real-time video model Lucy, on which Oasis 3 is built.
The Money And The Stack
Decart says the $300 million round was driven by what Leitersdorf calls huge demand increases for its e-commerce, live-streaming and physical-AI models. The lab credits much of its efficiency to DOS, the Decart Optimization Stack — vertically-integrated software that compresses inference cost across Nvidia, Amazon and Google hardware. Leitersdorf claims Decart has burned through drastically less than $100 million in its lifetime even while training frontier models.
Limits And Roadmap
TechCrunch's hands-on review found Oasis 3 produces the most photorealistic single-prompt environments of any commercial world model the publication has tested but flagged consistency drift: the system rapidly loses spatial integrity if you drive too far from the spawn point, and cars currently clip through other vehicles because the model has more good-driving data than crash data. Decart is now extending context length to store millions more tokens and plans a follow-up version that lets users seed simulations from a video instead of a single image. The startup expects developers to surface unforeseen use cases within months.
Where It Sits In The Race
Oasis 3 lands in a crowding field that already includes Google's Genie 3 research preview, Fei-Fei Li's World Labs Marble, and physics-aware video offerings from Luma and Runway. For more on the surrounding stack see NVIDIA and Hyundai's robotics alliance, RLWRLD's DexBench dexterity standard, and AGIBOT's Genie Envisioner 2 simulator.
Reporting based on coverage from TechCrunch, Dataconomy and Decart.