Back to Headlines
Tech
Jun 10, 2026
Analyzed by GPT OSS 120B

Decart’s Oasis 3 Brings Hours‑Long Photorealistic Driving Simulations, but Consistency Falters

AI Summary
Decart unveiled Oasis 3, a real‑time, photorealistic driving world model accessible via API and priced at $0.02 per second. While the model delivers unprecedented visual fidelity and long‑duration generation, it struggles with scene consistency and physics realism, highlighting early‑stage challenges in physical‑AI simulation.

Decart Launches Oasis 3, a Real‑Time Photorealistic Driving World Model

Decart announced Oasis 3 on 2026-06-10, offering developers an API that can generate immersive, multi‑camera driving environments for hours at a time. The startup positions the model toward autonomous‑vehicle firms and aims to grow a developer ecosystem similar to OpenAI’s language‑model API strategy.

Pricing, Funding, and Cost‑Efficiency Metrics

  • API usage: $0.02 per second (enterprise pricing varies by use case)
  • Recent financing: $300 million Series C, valuing Decart at ~$4 billion
  • Strategic investors: Toyota, Adobe, eBay, and existing backer Nvidia
  • Operational spend: under $100 million to date, thanks to the Decart Optimization Stack (DOS) that cuts compute costs by an order of magnitude

Implications for Autonomous‑Vehicle Development and the Wider Physical‑AI Ecosystem

Oasis 3’s ability to generate physically accurate, multi‑camera scenes on demand addresses a key bottleneck for training self‑driving systems: the scarcity of rare, edge‑case scenarios. By offering infinite generation rather than limited demos, Decart gives AV teams a scalable way to stress‑test perception and planning stacks. The model also pushes the broader world‑model market, competing with Google’s Genie 3, World Labs’ Marble, and video‑generation startups like Luma and Runway.

Technical Hurdles: Consistency Decay and Physics Gaps

Early testing reveals two major shortcomings:

  • Scene degradation: after a few minutes the generated environment drifts from the original prompt, turning a New York street into a generic Western city.
  • Physics realism: vehicles sometimes pass through one another, reflecting limited accident data in the training set.

The model’s auto‑regressive architecture—producing one frame (~8,000 tokens) at a time—fills its context window quickly, restricting long‑term memory. Decart is researching larger context windows and token compression to store “millions more tokens.”

Future Roadmap: Memory Extensions and a Developer‑Driven Surge

Leitersdorf predicts the next version will accept video seeds, improving continuity and potentially easing physics issues. More importantly, the company expects a rapid expansion of community‑built applications, echoing the early OpenAI API boom. In three months, Decart anticipates dozens of novel use cases emerging from its growing 100,000‑plus developer base.