Real-Time AI Video and World Models: The 2026 Release Tracker

Key takeaways

• ✓Two kinds of "real-time" ship today. World models generate an environment you move through (Project Genie, Odyssey-3, Oasis 3). Live video-to-video models transform a camera feed or a stream as it plays (PixVerse R1, LiveGen).
• ✓Most world models are still gated. Project Genie needs a US Google AI Ultra subscription, Runway Solaris is early access only, and Oasis 3 and Reactor are developer APIs. Odyssey-3's browser preview is the main one anyone can open for free.
• ✓Resolution has settled around 720p. Project Genie, PixVerse R1 (after its February update) and Odyssey-3 Pro all ship at or near 1280 × 720. Real-time generation still trades resolution for speed.
• ✓Speed tricks are converging. Runway, Odyssey and the TurboDiffusion team all describe the same recipe: generate frames autoregressively, distill sampling down to a few steps, then train the fast model on its own outputs.

Release table

Dates are announcement dates. Where a model was shown first and opened later, the notes below give both.

What changed in October 2026

Odyssey-3 opened to the public. Odyssey first unveiled Odyssey-3 on September 15, 2026, then opened a free research preview in October at experience.odyssey.systems, running the smaller Odyssey-3 Flash. You type a prompt, pick a first-person or third-person view, move through the generated world and trigger events. The base model has 14 billion parameters. Odyssey reports a Physics-IQ score of 66.1 for Odyssey-3 Pro, from a single run where one of eight generations was picked per task; without that selection it averaged 63.37 across four runs (The Decoder). Full entry: What is Odyssey-3?

Notes on each model

Runway GWM-1 (December 11, 2025). Runway's first general world model family, built on Gen-4.5, in three variants: GWM Worlds for explorable environments, GWM Avatars for conversational characters and GWM Robotics for manipulation. DeepLearning.AI reported output of up to two minutes at 1280 × 720 and 24 fps (Runway).

TurboDiffusion (December 23, 2025). Not a model you talk to but the reason many of these became fast. ShengShu Technology and Tsinghua University's TSAIL lab combined sparse attention, step distillation down to three or four sampling steps and 8-bit quantization. In their example, a 1080p, 8-second clip went from about 900 seconds to about 8 seconds on a single GPU (press release, code).

PixVerse R1 (January 12, 2026). A real-time world model built into PixVerse's consumer video app. The February 10 update added 720p output, built-in audio and limited partner API access; an April update added shared worlds, personalized avatars and removed the earlier five-minute session cap (PixVerse). Full entry: What is PixVerse R1?

Project Genie (January 29, 2026). Google DeepMind's experimental web app on top of Genie 3, with World Sketching, World Exploration and World Remixing modes. Sessions last 60 seconds at 720p and 24 fps, and access is limited to Google AI Ultra subscribers in the US aged 18 and over. Google has said it is a prototyping tool, not a game engine. Full entry: What is Genie 3?

Reactor (May 28, 2026). Reactor came out of stealth with $59 million in combined seed and Series A funding led by Lightspeed, with WndrCo and others. It sells a developer SDK and API for real-time generative video aimed at media, entertainment and robotics; its founders were technical leads on Apple Vision Pro (announcement).

Decart Oasis 3 (June 10, 2026). A world model that generates multi-camera driving scenarios from text prompts for training autonomous vehicles and robots, sold by API at $0.02 per second. It builds on Decart's real-time video model Lucy (Decart). Decart's earlier live video-to-video model is covered in What is MirageLSD?

Runway Solaris (August 31, 2026). The first "interface world model": instead of running code, it renders an app or website frame by frame as you use it, with a language model deciding what happens next and a real-time version of Gen-4.5 drawing it. In Runway's study of 250 people and nearly 7,500 judgments, Solaris was preferred over a coded interface built by Claude Opus 5 in 61% of comparisons for instruction following and 71% for natural behavior. Runway lists legible text, long-session coherence and screen-reader support as unsolved (Runway). Full entry: What is an interface world model?

LiveGen (our product). A consumer browser app for live camera transformation rather than world generation: it restyles you, your outfit or your whole scene while you talk and move, with no install. It sits in the video-to-video branch next to PixVerse R1, not in the world model branch. You can try it free.

How to read this table

• ✓"Real-time" means frames arrive while you act. A fast clip generator that returns a 10-second file in 10 seconds is not on this list; a model that keeps generating as you press keys or move on camera is.
• ✓Published specs are company claims. Resolution, frame rate and benchmark numbers come from each maker's announcement or the press coverage linked above, not from our testing.
• ✓Access changes fastest. Waitlists open, prices move and previews close. Check the linked source before planning around a row.