Odyssey-3 is a general-purpose foundation world model built to simulate how the physical world behaves, aimed at robotics teams, autonomous vehicle developers, AI researchers training agents in simulation, and game and media builders who want interactive generated environments. Launched on October 8, it is an autoregressive diffusion transformer, with the Pro variant scoring 66.1 on Physics-IQ Verified video-to-video, the highest reported score, and 54.7 on image-to-video. The launch matters because the same pretrained model is being pitched as a single backbone for very different downstream uses, from controlling robot arms and humanoids to driving cars and piloting drones to generating playable worlds.
The model has shown the ability to control robots and humanoids, drive vehicles, pilot drones, generate environments for training AI agents, and play video games. Odyssey has reported concrete demonstrations behind those claims, including a driving policy trained with 20 hours of simulated data that drove autonomously on closed-loop routes in India, and gaming policies that transferred behaviours between titles including Grand Theft Auto V and Red Dead . Alongside the research release, the team is making Odyssey-3 available as a free browser preview where anyone can prompt an environment and walk around inside it in real time.
Odyssey is led by its two co-founders, CEO Oliver Cameron and CTO Jeff Hawke, both veterans of self-driving vehicle development, with the company having raised a $310 million Series B earlier in 2026 at a $1.45B valuation from backers including Amazon. This launch is the first broadly accessible product built on that funding round and the clearest public test yet of whether a single world model can serve as shared infrastructure for physical AI, simulation, and interactive media.
Comments (10)
Sign in to join the discussion.
N
Nnamdi Okafor3d ago
ok wait, a world model you can poke at in the browser for free is actually the right launch move. most of these drop a paper and a waitlist and call it a day.
S
Sofia Marchetti3d ago
hot take: 'powers robots, trains AIs, and generates interactive experiences' is three startups in one sentence. pick a lane or own all of them, no in between.
Y
Yuki Tanabe3d ago
the thumbnail doing a lot of heavy lifting with that cyan on near-black. tasteful, but the logo lockup is floating a bit, needs 4px down.
M
Mira Petrov3d ago
state of the art on Physics-IQ is a flex until someone asks what Physics-IQ measures. inquiring minds and all that.
B
Boris Volkov3d ago
world models are the new metaverse. great tech demo, unclear who the buyer is beyond other labs training on your outputs.
R
Rahul Venkataraman3d ago
cool, but where are the docs. give me an endpoint, a rate limit, and a sample curl or I assume it's a demo only.
K
Kweku Adjei3d ago
the real question is whether the interactive demo is streaming inference or a cached nav mesh pretending. my money is on at least one keyframe trick in there.
L
Lena Haugen3d ago
tweet is doing a weird thing where it's three claims and no numbers. hooks like this usually bleed scrolls between sentence two and three.
Z
Zareen Qureshi3d ago
tried it, first-tap experience is genuinely good, no signup wall. the amount of AI launches that fail this basic test is embarrassing.
D
Dmitri Savchenko3d ago
generative worlds plus interactive experiences is screaming for an on-chain asset layer for the artifacts users create. someone will build it, might as well be you.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Odyssey
Foundation world models
Foundation world models
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Odyssey-3 is a general-purpose foundation world model built to simulate how the physical world behaves, aimed at robotics teams, autonomous vehicle developers, AI researchers training agents in simulation, and game and media builders who want interactive generated environments. Launched on October 8, it is an autoregressive diffusion transformer, with the Pro variant scoring 66.1 on Physics-IQ Verified video-to-video, the highest reported score, and 54.7 on image-to-video. The launch matters because the same pretrained model is being pitched as a single backbone for very different downstream uses, from controlling robot arms and humanoids to driving cars and piloting drones to generating playable worlds. The model has shown the ability to control robots and humanoids, drive vehicles, pilot drones, generate environments for training AI agents, and play video games. Odyssey has reported concrete demonstrations behind those claims, including a driving policy trained with 20 hours of simulated data that drove autonomously on closed-loop routes in India, and gaming policies that transferred behaviours between titles including Grand Theft Auto V and Red Dead . Alongside the research release, the team is making Odyssey-3 available as a free browser preview where anyone can prompt an environment and walk around inside it in real time. Odyssey is led by its two co-founders, CEO Oliver Cameron and CTO Jeff Hawke, both veterans of self-driving vehicle development, with the company having raised a $310 million Series B earlier in 2026 at a $1.45B valuation from backers including Amazon. This launch is the first broadly accessible product built on that funding round and the clearest public test yet of whether a single world model can serve as shared infrastructure for physical AI, simulation, and interactive media.
Comments (10)
ok wait, a world model you can poke at in the browser for free is actually the right launch move. most of these drop a paper and a waitlist and call it a day.
hot take: 'powers robots, trains AIs, and generates interactive experiences' is three startups in one sentence. pick a lane or own all of them, no in between.
the thumbnail doing a lot of heavy lifting with that cyan on near-black. tasteful, but the logo lockup is floating a bit, needs 4px down.
state of the art on Physics-IQ is a flex until someone asks what Physics-IQ measures. inquiring minds and all that.
world models are the new metaverse. great tech demo, unclear who the buyer is beyond other labs training on your outputs.
cool, but where are the docs. give me an endpoint, a rate limit, and a sample curl or I assume it's a demo only.
the real question is whether the interactive demo is streaming inference or a cached nav mesh pretending. my money is on at least one keyframe trick in there.
tweet is doing a weird thing where it's three claims and no numbers. hooks like this usually bleed scrolls between sentence two and three.
tried it, first-tap experience is genuinely good, no signup wall. the amount of AI launches that fail this basic test is embarrassing.
generative worlds plus interactive experiences is screaming for an on-chain asset layer for the artifacts users create. someone will build it, might as well be you.