Decart’s Oasis 3 serves up photorealistic driving environments at $0.02 per second and claims to run infinitely—but push the simulation for more than a few minutes and the world forgets you were ever there.
Oasis 3 is the latest interactive world model from the startup behind the real-time video model Lucy. Available via API starting Wednesday, it targets autonomous vehicle companies that need to spin up rare driving scenarios at scale. Decart also wants developers to build on top of it, betting that a world-model platform can replicate OpenAI’s language-model ecosystem play. More than 100,000 developers already use Lucy; Oasis 3 is built on that foundation.
Oasis 3 Cheaper by an Order of Magnitude
Pricing at two cents per second undercuts rivals by a lot. CEO Dean Leitersdorf credits the Decart Optimization Stack (DOS), which squeezes models onto Nvidia, Amazon, and Google hardware. “By being so vertically integrated, we’re able to be more than an order of magnitude cheaper than anyone else in the industry in order to run these models.” The startup has burned through “drastically less” than $100 million lifetime, he says.
Decart raised $300 million a few weeks before this launch, hitting a $4 billion valuation. Strategic investors include Toyota, Adobe, eBay, and existing backer Nvidia—all potential customers. The company is playing in a crowded arena: Google’s Genie 3, Fei-Fei Li’s World Labs Marble, and video generation startups like Luma and Runway are all chasing world models for physical AI.
Auto-Regressive Frame Generation Breaks Consistency
Oasis 3 generates multi-camera environments (one front-facing, two side-facing) from a single text prompt. Initial scenes are photorealistic and beautiful. But the model is auto-regressive: it produces one frame at a time, looking back at previous frames to decide what comes next. Each frame is roughly 8,000 tokens. At tens of frames per second, the context window fills up fast. “We’re researching how to do longer context to store millions more tokens, and how to compress the memory into fewer tokens,” Leitersdorf says.
In testing, the thematic integrity degraded rapidly. A prompted New York City morning scene turned into a generic urban street within minutes. Turning around to the original intersection? Gone, replaced by an entirely new environment. Controls become unresponsive; the car drives through other vehicles—the model doesn’t simulate contact physics properly. Leitersdorf calls that a “major research problem,” noting that “there’s drastically more data on good driving compared to accidents.”
Leitersdorf believes the next version, which will accept a video as input instead of a single image, may partially solve consistency. Until then, Oasis 3 is a stunning proof-of-concept that can generate hours of disjointed, dreamlike driving—useful for variety but not yet a reliable physics simulator.
Decart’s bet rests on developers treating this as a programmable world model, not a finished simulation. If they can compress memory and extend context, the $0.02-per-second pipeline could become the default sandbox for physical AI training.
Source: Decart's new world model can simulate hours of photorealistic driving - with some caveats
Domain: techcrunch.com
Comments load interactively on the live page.