Decart has opened access to Oasis 3, a ”world model” that can generate photorealistic virtual road scenes in real time and expose them through an API from day one. The startup is aiming first at autonomous driving, where companies need millions of test scenarios – including the rare, ugly ones that are almost impossible to capture on actual roads – and then at robotics and other forms of physical AI.
Decart says Oasis 3 costs 2 cents per second of generation, with enterprise pricing handled separately. The API is the whole point: the startup wants developers to create endless driving environments without building a real city block every time they want to test a model.
The execution is less simple, because this is one of the hottest corners of AI right now, with Google, World Labs, Luma, and Runway all pushing their own takes on generated worlds and simulations.
Oasis 3 pricing and API access
Decart says Oasis 3 costs 2 cents per second of generation, with enterprise pricing handled separately. That matters because the company is not just selling a demo; it is trying to turn world generation into a developer platform.
The startup already says it has a community of more than 100,000 developers, many of them using its earlier Lucy model for e-commerce and live-streaming use cases. That gives Decart something many model makers envy: a base of people who might actually build on top of the thing instead of just applauding it on a launch page.
How Oasis 3 generates driving scenes
According to Decart, Oasis 3 creates physically plausible environments from three cameras at once: a front view plus two side views. The company says the system can generate road situations almost without limit, which is exactly what autonomy teams want when they are hunting for weird edge cases rather than normal traffic.
The technical trick sits inside Decart’s own DOS, short for Decart Optimization Stack, which the company says helps run models on Nvidia, Amazon, and Google hardware. Decart claims that setup lets it operate world models at more than an order of magnitude lower cost than competitors. If true, that is the kind of advantage that can turn a flashy research demo into something infrastructure buyers actually budget for.
Where Oasis 3 still falls apart
For all the hype, Oasis 3 is not a solved problem wearing a nicer demo. During testing, the system could generate convincing scenes from text prompts, but it gradually lost consistency: a New York street could mutate into a generic cityscape, and intersections could disappear after a user drove away and came back.
Physics is still a mess too. Cars can drive through other cars, because the model has trouble keeping objects coherent over time, which is a rather awkward flaw for something pitched to autonomous-vehicle developers. The reason is baked into the architecture: Oasis 3 is autoregressive, so it generates frames one after another and keeps leaning on what it already made.
- Each frame contains about 8,000 tokens of information.
- Generating dozens of frames per second burns through context fast.
- Decart is working on more memory and better compression.
Toyota, Adobe, eBay and Nvidia are already in
Decart’s launch comes just weeks after the two-year-old startup raised $300 million and hit a valuation of almost $4 billion, with Toyota, Adobe, eBay, and Nvidia among the investors. That list is telling: the money is not only coming from people who want prettier simulations, but from companies with real reasons to care about commerce, content, and compute.
The company says demand has surged from e-commerce, streaming services, and physical AI projects. It also says the next version of Oasis will be able to build worlds not only from images but from video of real locations, which could make the scenes stick together better. The bigger question is whether Decart can turn a clever API into a platform before rivals copy the idea and the novelty wears off.