Rare crash scenarios used to cost AV teams months of physical road time
Autonomous vehicle companies spend enormous resources recreating edge-case driving scenarios — black ice at dusk, debris mid-lane, sensor-confusing sun glare — because real-world data collection for rare events is slow, dangerous, and expensive. There is no clean way to run those scenarios at scale without a simulation layer that actually looks like the real world.
A world model that runs like a video game, billed like an API
Decart’s new world model can simulate hours of photorealistic driving — with some caveats takes a prompt or programmatic input via API and outputs photorealistic, interactive driving environments that generate in real time and run indefinitely without hitting a time wall. You call the API, specify the scenario parameters, and get back a navigable world your software can drive through, log, and stress-test. The model is built on Decart’s existing real-time video foundation, which means it inherits physics-aware rendering rather than stitching together pre-baked clips.
AV engineers are the first ones who should be watching this
- Simulation engineers at autonomous vehicle companies who need thousands of hours of rare-event driving data without sending a test fleet into a snowstorm
- Robotics researchers who need photorealistic physical environments to train manipulation and navigation models without building physical test rigs
- Developer-side AI builders who want a programmable world layer the way they once wanted a programmable language layer from OpenAI’s API
The developer angle is the one worth watching long term: Decart already has 100,000 developers building on its video model Lucy, and Oasis 3 is the direct extension of that community into physical AI.
The world model market just got its first commercial API with a per-second price
Google’s Genie 3 is still in research preview, World Labs launched Marble for commercial use cases, and Luma and Runway are moving their video physics into simulation territory, but none have shipped a public API with transparent per-second pricing. At a $4 billion valuation and with Toyota, Adobe, and Nvidia on the cap table, Oasis 3 is positioned as infrastructure rather than a demo, and the companies that build on it earliest will have a data advantage that compounds.
What you can actually do with it today
- Generate hours of rare driving scenarios — black ice, debris, edge-case intersections — on demand
- Stress-test AV software against photorealistic edge cases without physical road access
- Build programmatic simulation pipelines using the API at $0.02 per second of generated world
- Prototype physical AI applications in robotics using real-time interactive environments
Pricing starts at $0.02 per second of generated simulation, with enterprise tiers negotiated by use case.
The caveat in the headline is real
Photorealism and infinite generation are the headline, but long-horizon physical consistency — whether a simulated world holds its geometry and causality across extended complex interactions — remains an open research problem for every world model on the market, and Oasis 3 is not exempt from that ceiling.
Alternatives worth knowing
World Labs’ Marble targets similar commercial simulation use cases and is worth a direct comparison for enterprise buyers. For teams already inside a video generation workflow, Luma’s physics-aware models are the closest adjacent option, though neither ships with the same per-second API pricing model.
World models are becoming the new API layer for physical AI
The shift from language models as the default API primitive to world models is happening faster than most teams have planned for. We cover tools like this every Friday — subscribe here and we’ll send the best ones straight to you.