Decart unveils Oasis 3 world model for photorealistic driving simulations

World models like Oasis 3 are increasingly seen as a missing piece between today’s pattern‑matching LLMs and systems that can reason about and act in the physical world. Decart is positioning Oasis 3 as a programmable, photorealistic driving simulator that developers can call over an API, effectively turning high‑fidelity synthetic data into an on‑demand service. Combined with its highly optimized DOS stack, the company claims it can generate these environments cheaply enough to be used at scale, not just in research labs.([techcrunch.com](https://techcrunch.com/2026/06/10/decarts-new-world-model-can-simulate-hours-of-photorealistic-driving-with-some-caveats/))

From an AGI race lens, this matters because robust world models are one plausible path to general cognition. If you can train agents across billions of varied, physically‑plausible scenarios, you start to approximate the kind of rich experience base humans accumulate in the real world—but safely and at machine speed. Oasis 3 isn’t there yet; TechCrunch notes visible degradation over time and physics inconsistencies like cars passing through each other. But it’s a concrete step toward interactive, persistent environments where agents can explore, plan, and learn complex behaviours.

Strategically, Decart is also trying to own the developer ecosystem for world models, mirroring how OpenAI built an API‑centric platform for text models. If it succeeds, other world‑model efforts at Google, NVIDIA, and startups like Luma or Runway will feel pressure to match both fidelity and cost. That competition is likely to accelerate progress on physically grounded models—one of the key ingredients for embodied and agentic systems that start to look a lot more like proto‑AGI than today’s chatbots.