Decart Oasis 3 API Renders Endless Driving Sims at 22 FPS

Decart has shifted its focus from general generative video to specialized infrastructure with the launch of Oasis 3, a real-time interactive world model for autonomous vehicle validation. Previous generative video tools produced static, fixed-length outputs. Oasis 3 functions as a closed-loop environment that responds directly to API commands and real-time control signals. The system is available immediately to developers via API, targeting edge-case generation for driving agents.

Simulation Performance and Architecture

Oasis 3 outputs interactive 3D environments at a throughput of 22 frames per second (fps) at a resolution of 768px. The model maintains an end-to-end latency of under 200ms, which provides the immediate visual feedback required by steering and acceleration inputs. For hardware training setups, the system generates synchronized multi-camera views, such as simultaneous forward and side-facing feeds, extracted from a single internal world state.

The underlying architecture relies on DOS 2.0 (Decart Optimization Stack). This proprietary training and serving platform optimizes the workload across NVIDIA, Amazon, and Google accelerators. By standardizing AI inference at the stack level, Decart states that execution costs are reduced by more than an order of magnitude compared to standard deployment pipelines.

Learned Physics vs Traditional Engines

Autonomous vehicle validation traditionally relies on hand-built physics engines like CARLA or NVIDIA Isaac Sim. Oasis 3 replaces rigid programmatic physics with a learned foundation model, sharing architectural similarities with recent AI world simulation pipelines. This approach allows for endless, dynamic scenario generation without manual 3D asset creation.

The tradeoff for generative diversity is physical consistency. Researchers have observed intermittent environmental degradation over long simulation durations. The model also occasionally hallucinates physical collisions, resulting in “ghosting” where objects pass through one another. If you build safety-critical validation pipelines, these violations of rigid physical laws require secondary filtering before you can use the synthetic data for compliance testing.

Pricing and Commercial Scale

Developers can access the Oasis 3 API at a standard rate of $0.02 per second of generated simulation, with custom enterprise tiers available. The API launch is backed by a $300 million funding round from May 2026 led by Radical Ventures, with participation from Nvidia, Toyota Ventures, Sequoia, and Benchmark. This round valued Decart at approximately $4 billion. The company reports a cumulative cash burn of less than $100 million, attributing the capital efficiency to the hardware utilization rates achieved by the DOS 2.0 stack.

When designing multi-agent systems for autonomous driving, treat generative world models as a supplement for rare edge cases rather than a complete replacement for deterministic physics simulators. The API is best utilized for generating unpredictable hazardous scenarios that are too dangerous to capture on real roads, provided your testing pipeline automatically discards runs containing physics hallucinations.