Odyssey-3 World Model Reports 66.1 on Physics-IQ and Opens a Free Research Preview
Odyssey-3 Pro scored 66.1 on the video-to-video track of Physics-IQ Verified, according to the lab's October 8 launch post. Odyssey's footnotes attach conditions to that number. It is a best-of-eight result with prompt enhancement, generated at 1280x720. Pro with prompt enhancement and a single sample scored 63.4. The base Odyssey-3 at 832x480 scored 51.8 on unmodified prompts, 61.6 with prompt enhancement and 64.4 at best-of-eight. On the image-to-video track, Pro reached 54.7.
Physics-IQ and WorldMark
Physics-IQ Verified, developed by Anates Labs and Google DeepMind, asks a model to continue real recordings of physical experiments in fluid dynamics, optics, solid mechanics, magnetism and thermodynamics. Odyssey's comparison set includes FLUX 3, Cosmos3, Sora 2 and Veo 3.1, using leaderboard data as of October 7. Its cost figures assume $1 per MI355X GPU-hour.
On WorldMark, which averages 13 metrics covering control-following, visual quality and world memory, Odyssey-3 ranked first in three of four splits. In first-person stylized scenes it scored 77.2 against 77.0 for LingBot-World. In third-person real scenes it scored 79.0 against 76.9 for HY-World 1.5. In first-person real scenes it placed third at 80.6, behind Lyra 2.0 at 84.4 and AlayaWorld at 83.0.
The model is an autoregressive video diffusion transformer trained with teacher forcing and causal masking, then distilled to a few-step sampler for real-time use. Its training data comes from internet video with time-localized event annotations, gameplay recordings paired with keyboard and mouse input, and simulated rigid-body interactions. A free research preview runs at experience.odyssey.systems, and developers can request API access at developer.odyssey.ml
Google's Genie 4, covered in our earlier piece, Genie 4 Adds Native Audio and a Planning Model to Real-Time World Simulation, also arrived without latency or session-length figures. Odyssey's post supplies benchmark tables and leaves the frame rate unstated.
Policies for robots and cars
An earlier Odyssey post, from September 15 and bylined to Oliver Cameron and Jeff Hawke, describes using the frozen world model as a base for control. A learned action decoder translates the model's internal representations into commands for each machine. Robot arms completed manipulation tasks after tens of hours of demonstrations, including regrasping after a missed grasp, a behavior absent from the training data. Flexion, a humanoid robotics company, built whole-body control policies on the model from tens of hours of teleoperation data. Odyssey reports they kept working under lighting changes that broke the vision-language-action baselines tested, without giving numeric scores.
The driving result carries the only comparison with a number attached. A waypoint-predicting policy trained on 20 hours of simulated driving drove in closed loop on streets in India. Against policies trained on real driving footage, the simulation-trained versions covered about 77% as much distance between safety-driver interventions. The drone work took place in a simulated indoor space. Odyssey calls its results early and writes that frontier world models sit roughly two orders of magnitude behind language models in scale.
Company and prior research
Cameron, the CEO, co-founded Voyage, which Cruise acquired, and later served as Cruise's VP of product. Hawke, the CTO, was an engineer at Wayve. They founded Odyssey in 2023. The June 17 Series B was led by Natural Capital, with Amazon, AMD Ventures, GV, EQT and In-Q-Tel participating. TechCrunch put total funding at $337 million. AWS is the preferred cloud provider, and Odyssey is working with Amazon's Annapurna Labs on Trainium chips. The company has not disclosed revenue or customers.
The lab's output over the past year has included several papers and models. Odyssey-2 streams a new frame every 50 milliseconds, or 20 frames per second. The Starchild-1 technical report describes joint audio and video generation at up to 24 fps. Against the bidirectional Ovi model it was distilled from, imaging quality rose from 0.451 to 0.561 while the CLAP audio-text score fell from 0.265 to 0.182, on clips of four to five seconds. Agora-1 renders a shared four-player GoldenEye match.
PROWL-1 (arXiv 2605.18803) sends a reinforcement-learning agent through MineRL to find a world model's failures and feeds them back as training data. PROWL-2 extends this to multiple agents with a fidelity gate that sends unreliable imagined experience to repair. It reports the highest mean win rate in all nine SMACv2 scenarios, and on the Gate-3 quadruped task it reaches 70.4% success against 29.8% for the SeqWM baseline. All authors are Odyssey staff or affiliated academics, and the PROWL-2 results use budgets that differ from the original backbone papers.
The launch page lists no product pricing. Its only cost figures are benchmark estimates built on the assumed $1 per GPU-hour.
Author
David Borish writes about frontier AI and is the author of The Tony Hawk Paradox: When Video Games Predict Reality. More at davidborish.com
