--- title: Odyssey-3 launches as a world model topping the... description: Odyssey released Odyssey-3, with its Odyssey 3 Pro variant claimed to top the Physics-IQ Verified benchmark among video-to-video models for physical accuracy.... canonical: https://daily.dev/posts/odyssey-3-launches-as-a-world-model-topping-the-physics-iq-verified-benchmark-o5ogrscrs twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Odyssey-3 launches as a world model topping the Physics-IQ Verified benchmark | daily.dev og:description: Odyssey released Odyssey-3, with its Odyssey 3 Pro variant claimed to top the Physics-IQ Verified benchmark among video-to-video models for physical accuracy.... og:url: https://daily.dev/posts/odyssey-3-launches-as-a-world-model-topping-the-physics-iq-verified-benchmark-o5ogrscrs og:image: https://api.daily.dev/og/posts/O5oGRScrs.png og:image:alt: Odyssey-3 launches as a world model topping the Physics-IQ Verified benchmark og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Odyssey-3 launches as a world model topping the Physics-IQ Verified benchmark **[Collections](https://daily.dev/sources/collections)** · 3 min read · 0 upvotes · 0 comments ## Summary Odyssey released Odyssey-3, with its Odyssey 3 Pro variant claimed to top the Physics-IQ Verified benchmark among video-to-video models for physical accuracy. Unlike a standard video generator, the world model reportedly learns how physical systems work from visual observations and maps that understanding to control systems, enabling it to control robots, drive cars, play video games, or pilot drones. It is currently available only as a research preview, and the benchmark claim comes from Odyssey and a third-party reading of the leaderboard rather than independent verification. ## Content Odyssey has released Odyssey-3, a research preview of its most powerful foundation world model. The Palo Alto startup says one underlying model can control robot arms and humanoids, drive a car on real roads in India, fly drones indoors, and play Grand Theft Auto V. ## How it works Odyssey-3 is an autoregressive diffusion transformer. It predicts the next frames from earlier ones plus incoming actions, so a prompted scene reacts to each move. Give it a prompt and it builds an environment, then predicts changes in real time as a person walks through it or triggers events. Users can move in first or third person, and the camera can also move independently. There are two versions. Odyssey-3 Flash is a distilled few-step model that turns a text prompt into an explorable world. Anyone can try it free today. Odyssey-3 Pro is the larger model behind the benchmark numbers. ## Benchmarks Odyssey-3 Pro scored 66.1 on Physics-IQ Verified video-to-video with best-of-8 sampling, the highest reported result. The benchmark asks models to continue videos of real physics experiments. On image-to-video, Pro roughly ties FLUX 3 [large]. Odyssey also claims first place in three of four WorldMark categories, based on its own evaluation. ## Robotics and driving This is the part that stood out to me. Odyssey adapted the model to robot arms, humanoids, and a car, and each system uses its own trained policy on top of Odyssey-3. - With Odyssey-3 frozen, a policy trained on 20 hours of Indian road data drove a real car. - With tens of hours of demos, a robot arm recovered from a missed grasp, a behavior that never appeared in those demos. - Humanoid robotics company Flexion built control policies on top of the model. Neither the driving nor the grasping result comes with reported success rates, so it's hard to say how reliable they are. Recovering from a failure it never saw is still the kind of thing you'd want from a world model. ## The field Odyssey raised $310 million in June at a $1.45 billion valuation. It's entering a crowded space: AMD acquired World Labs for $8.2bn, startup Emulate is reportedly in talks to raise $700m, and Nvidia has released its open Cosmos models. Some commentators are already predicting that every robot, humanoid and self-driving car will run on a world model like this within a few years. That's a big leap from a research preview with a handful of demos. But a single model that handles a car, an arm and a game is a real step. ## Questions this post answers ### What is Odyssey 3 Pro and how is it different from a regular video generation model? Odyssey 3 Pro is a world model that goes beyond generating video: it takes visual observations from the world, learns how physical systems behave, and maps that knowledge onto physical controls. This lets the same model control a robot, drive a car, play a video game, or pilot a drone, rather than just producing video output. _daily.dev surfaces emerging world model releases like this for engineers tracking generative AI beyond video generation._ ### Does Odyssey-3 actually rank first on the Physics-IQ Verified benchmark? Odyssey 3 Pro is reported to rank number one among video-to-video models on the Physics-IQ Verified benchmark for physical accuracy, according to Odyssey's own claim and an independent reading of the leaderboard. The model is only available as a research preview, so the ranking has not been independently confirmed beyond that reporting. _Developers evaluating world model benchmarks can follow claims like this on daily.dev before trusting vendor leaderboards._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 15 comments across x (as of 2026-10-09). **TL;DR:** People are intrigued by the generative-world-plus-robotics combo, especially the unscripted grasp recovery and the low-data driving policy, but several push back on how much the benchmark number and the 'generalization' claim actually prove. **Sentiment:** 40% positive · 45% mixed · 15% skeptical **The case for** - The action-reaction loop (generate a world, interact, predict what happens) is seen as a meaningfully different direction than pure pixel generation. - Recovering from a missed grasp that wasn't in training data is called out as a genuinely impressive, screenshot-worthy moment. - The India driving policy reportedly trained on just 20 hours of data with a frozen world model, suggesting the robotics bottleneck could shift from data collection to backbone selection. - Collecting only tens of hours of demonstration data is seen as a low bar relative to the physics results achieved. **The pushback** - The headline 66.1 score comes from best-of-8 sampling, and some want to know what a single-attempt score looks like. - In the first-person-real split specifically, the model reportedly ranks third behind two other models, undercutting the single headline number. - Some argue the grasp recovery looks like interpolation within the demonstration manifold rather than evidence of new physics understanding, and may break under real-world contact slips. - A world generated per-prompt without exact reproducibility is viewed as unsuitable for games that need the same level replayed identically. **By community** - x (mixed): Reactions mix genuine enthusiasm for the robotics and physics results with pointed questions about benchmark methodology and how far the generalization claims really extend. **Hottest debate:** Whether the missed-grasp recovery demonstrates real generalization beyond training data or is just interpolation within the demo set that won't hold up outside controlled conditions. **Open questions** - What is the single-sample (non best-of-8) Physics-IQ score? - What did the benchmark's wrong answers actually look like, especially at physics edge cases? - Does the grasp-recovery behavior hold up with real-world contact slips outside the lab setup? **Highlights** > @testingcatalog 66.1 is with best-of-8 sampling tho. whats the score on a single try? thats what users actually get > — [cortexch on x](https://x.com/cortexch/status/2108261357582008798) > @omarsar0 @odysseyml recovery looks like interpolation inside the demo manifold, not new physics. contact slips usually break that outside the lab setup. > — [okleusi on x](https://x.com/okleusi/status/2108252944462475390) > @testingcatalog in first person real it sits third, behind Lyra 2.0 and AlayaWorld, but it leads the other three splits. that spread says more than the one 66.1 headline > — [fireplyai on x · 1 points](https://x.com/fireplyai/status/2108247759249695184) > @testingcatalog the number i'd underline is further down odyssey's page: the india driving policy was trained on 20 hours of driving data, with the world model frozen. if that holds up, the robotics bottleneck shifts from collecting data to picking the backbone. > — [maxxingtokens on x](https://x.com/maxxingtokens/status/2108269820160463296) > @testingcatalog Great for walking around, but a game needs the same world twice. My roguelike goes the other way: every tile is pixel art generated in code, and the same seed gives the same map. Until a world model can replay a level exactly, it's a demo engine, not a level engine. > — [jpcardama on x](https://x.com/jpcardama/status/2108259351488282695) **Source threads** - [x](https://x.com/testingcatalog/status/2108246975883751665) · 1 points · 8 comments - [x](https://x.com/omarsar0/status/2108245674961588672) · 0 points · 7 comments ## Similar posts on daily.dev - [World model maker Odyssey nabs $1.45B valuation backed by Amazon and other big names](https://daily.dev/posts/world-model-maker-odyssey-nabs-1-45b-valuation-backed-by-amazon-and-other-big-names-wt1cfwtg6) · TechCrunch · 0 upvotes · 0 comments - [The Waymo World Model: A New Frontier For Autonomous Driving Simulation](https://daily.dev/posts/the-waymo-world-model-a-new-frontier-for-autonomous-driving-simulation-5jsu87osj) · Hacker News · 0 upvotes · 0 comments - [Into the Omniverse: Physical AI Open Models and Frameworks Advance Robots and Autonomous Systems](https://daily.dev/posts/into-the-omniverse-physical-ai-open-models-and-frameworks-advance-robots-and-autonomous-systems-dtpus18ht) · NVIDIA · 0 upvotes · 0 comments --- Tags: [#genai](https://daily.dev/tags/genai), [#robotics](https://daily.dev/tags/robotics) [View this post on daily.dev](https://daily.dev/posts/odyssey-3-launches-as-a-world-model-topping-the-physics-iq-verified-benchmark-o5ogrscrs) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Odyssey-3 launches as a world model topping the Physics-IQ Verified benchmark","url":"https://daily.dev/posts/odyssey-3-launches-as-a-world-model-topping-the-physics-iq-verified-benchmark-o5ogrscrs","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/odyssey-3-launches-as-a-world-model-topping-the-physics-iq-verified-benchmark-o5ogrscrs"},"datePublished":"2026-10-08T16:58:29.642Z","dateModified":"2026-10-09T05:01:55.648Z","description":"Odyssey released Odyssey-3, with its Odyssey 3 Pro variant claimed to top the Physics-IQ Verified benchmark among video-to-video models for physical accuracy....","image":"https://pbs.twimg.com/media/HUH5l8Sb0AE8FVD.jpg","thumbnailUrl":"https://pbs.twimg.com/media/HUH5l8Sb0AE8FVD.jpg","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/odyssey-3-launches-as-a-world-model-topping-the-physics-iq-verified-benchmark-o5ogrscrs","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"genai,robotics","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Odyssey-3 launches as a world model topping the Physics-IQ Verified benchmark"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/odyssey-3-launches-as-a-world-model-topping-the-physics-iq-verified-benchmark-o5ogrscrs#faq","mainEntity":[{"@type":"Question","name":"What is Odyssey 3 Pro and how is it different from a regular video generation model?","acceptedAnswer":{"@type":"Answer","text":"Odyssey 3 Pro is a world model that goes beyond generating video: it takes visual observations from the world, learns how physical systems behave, and maps that knowledge onto physical controls. This lets the same model control a robot, drive a car, play a video game, or pilot a drone, rather than just producing video output. daily.dev surfaces emerging world model releases like this for engineers tracking generative AI beyond video generation."}},{"@type":"Question","name":"Does Odyssey-3 actually rank first on the Physics-IQ Verified benchmark?","acceptedAnswer":{"@type":"Answer","text":"Odyssey 3 Pro is reported to rank number one among video-to-video models on the Physics-IQ Verified benchmark for physical accuracy, according to Odyssey's own claim and an independent reading of the leaderboard. The model is only available as a research preview, so the ranking has not been independently confirmed beyond that reporting. Developers evaluating world model benchmarks can follow claims like this on daily.dev before trusting vendor leaderboards."}}]} ```