The demo begins with a crisp, high-definition view of a New York City street at dawn. It looks indistinguishable from a dashcam recording. Then, you start driving. Within minutes, the city dissolves into a generic urban blur, and the intersection you just passed has vanished entirely.

This is the promise and the peril of Oasis 3, the latest "world model" from AI startup Decart. Released Wednesday, the model aims to solve a massive bottleneck in autonomous vehicle development: the need for infinite, diverse training data. By generating photorealistic, interactive environments in real time, Decart hopes to move beyond static datasets and into the realm of programmable, simulated reality.

The Efficiency Play

Decart is betting that its secret sauce isn't just the model—it’s the plumbing. The company has built a proprietary optimization stack, the DOS, designed to squeeze maximum performance out of Nvidia, Amazon, and Google hardware.

CEO Dean Leitersdorf claims this vertical integration makes Oasis 3 more than an order of magnitude cheaper to run than competing systems. At $0.02 per second, the cost is low enough to encourage developers to treat the model like a sandbox. It is a deliberate strategy. By opening API access from day one, Decart is attempting to replicate the developer-first ecosystem that propelled OpenAI to dominance.

Where the Simulation Breaks Down

Despite the technical efficiency, the experience of using Oasis 3 reveals the limitations inherent in current generative AI. While the initial frames are stunning, the model struggles with object permanence and thematic consistency.

If you drive for too long, the world begins to drift. A specific New York street might morph into a generic Western city. Turn the car around, and the road you just traveled is gone, replaced by an entirely new, procedurally generated landscape. It feels less like a rigid simulation and more like a fever dream.

Physics, too, remains a significant hurdle. In my testing, the car frequently clipped through other vehicles as if they were ghosts. Leitersdorf acknowledges this is a "major research problem." The issue, he explains, is a data imbalance: there is an abundance of footage showing perfect driving, but very little data on how to simulate the chaotic, messy reality of a collision.

What This Means for Developers

For autonomous vehicle companies, the value proposition is clear: they need to test edge cases without putting real cars on the road. Oasis 3 allows for the creation of multi-camera environments—one front-facing and two side-facing—that can be generated infinitely.

However, developers should be wary of the model's current "hallucination" rate. If a simulation cannot maintain the integrity of a stop sign or the physical boundaries of a lane, its utility for safety-critical training is limited.

Key Takeaways

  • Infinite Generation: Oasis 3 allows for continuous, real-time environment generation, moving away from the limited, pre-rendered clips used by older simulators.
  • Hardware Optimization: Decart’s proprietary stack allows it to run models at a fraction of the cost of competitors, targeting a $0.02 per second price point.
  • The Physics Gap: The model currently struggles with object permanence and collision physics, meaning it is not yet a replacement for high-fidelity physics engines.

Decart has raised $300 million to date, valuing the company at nearly $4 billion. With backing from Toyota and Nvidia, they have the capital to solve these research problems. The question is whether they can fix the underlying physics before their competitors do. The next few months of API usage will tell us if this is a viable tool for engineers or just a very expensive, very pretty toy.