Odyssey-2 is a family of real-time, interactive generative models built by the AI lab Odyssey. The family has four members: an undated Odyssey-2 Alpha, Odyssey-2 (October 2025), Odyssey-2 Pro (January 2026) and Odyssey-2 Max (April 2026). It follows two earlier models from the same lab, Explorer (December 2024) and Odyssey-1 (May 2025), and precedes Odyssey-3, announced on 15 September 2026. [1][4][11][13][15]

The line traces an unusually visible change of direction. Explorer was a still-image-to-3D tool aimed at film and game production, built on Gaussian splatting. Odyssey-1 was a navigable, streaming video demo. By Odyssey-2 Pro the lab was selling a developer API, and by Odyssey-2 Max it was publishing physics benchmark tables against NVIDIA Cosmos and pitching robotics, defence and simulation customers. [1][4][13][15] Odyssey also rewrote the earlier announcements to match the later framing: the May 2025 launch was published as "interactive video" and was retitled "Odyssey-1: A Playable World Model" roughly nine months later. [5][4]

Lineage

Model	Announced	Announcement	What was new	Availability at launch
Explorer	18 December 2024	"World Models for Film, Gaming, and Beyond" (current title; originally "Generative World Models for Film, Gaming, and Beyond")	Image-to-world generation output as Gaussian splats, editable in standard 3D tools	Seeded to production houses and artists; applications via the blog [2]
Odyssey-1	28 May 2025	Published as an interactive video research preview; retitled "Introducing Odyssey-1: A Playable World Model" in early 2026	Real-time navigable video stream, a new frame every 40 ms	Free public research preview on the web [6]
Odyssey-2 Alpha	Not announced in a dedicated post	Listed only on the research index	Described by Odyssey as adding generality to Odyssey-1's navigation input	Not stated by Odyssey [9][10]
Odyssey-2	27 October 2025	"Introducing Odyssey-2"	Open-ended text prompting during a running stream; 20 FPS	Free web demo; API announced as "coming soon" [11][12]
Odyssey-2 Pro	23 January 2026	"The GPT-2 Moment for World Models Is Here"	Larger model, 720p at 22 FPS; first Odyssey model in an API	Public developer API with JavaScript and Python SDKs [13]
Odyssey-2 Max	21 April 2026	"Introducing Odyssey-2 Max: Scaled World Simulation"	Roughly 3x the parameters and 10x the training compute of Odyssey-2 Pro	Private beta for partners [15]

The Odyssey-1 row needs a caveat. The 28 May 2025 post now carries the title "Introducing Odyssey-1: A Playable World Model", but the page archived at the original odyssey.ml domain as late as 21 January 2026 was titled "AI video you can both watch and interact with in real-time" and called the system an interactive video model, not a world model. [5][4]

Before Odyssey-2: Explorer

Odyssey announced Explorer on 18 December 2024, in the same post that announced Pixar co-founder Ed Catmull joining the board of directors and investing in the company. [1][2] The post frames the work squarely as a creative-tools play: Odyssey argued that hand-building worlds in 3D software is both the enabler and the bottleneck of film and games, and positioned Explorer as a way to shorten that work. [1]

Explorer was an image-to-world model. It took an image (TechCrunch reported it also accepted text captions such as "a Japanese garden, with rich, green foliage") and produced a 3D scene represented as Gaussian splats. [1][2] Odyssey said it chose splats because the representation had become the focus of graphics and vision research and because creative tools were adding support for it, so a generated world could be opened in Unreal, Houdini, Blender, Maya, 3D Studio Max or After Effects and edited by hand. [1]

Odyssey published concrete limitations at the time. A generation took an average of 10 minutes, and the company listed resolution and world completeness as active work, along with better controllability through video-to-world and world-to-world inputs. [1] TechCrunch added that scenes were low resolution and carried visual artifacts. [2] Odyssey said it had tested Explorer on the virtual production stage at Garden Studios in London and had seeded the model to production houses and independent artists who could apply through the blog. [1][2]

The training data behind Explorer was the company's other unusual bet. In a post dated 13 November 2024 announcing an $18 million Series A led by EQT Ventures, with GV and Air Street Capital participating, Odyssey described a backpack-mounted capture rig built with the optical imaging company Mosaic: roughly 25 pounds, with 6 cameras, 2 lidars and an inertial measurement unit, capturing 360 degree scenes at 13.5K resolution with depth. [3] The team's stated rationale was that its founders and more than 90% of its technical staff had come from self-driving programmes at Cruise, Wayve, Waymo and Tesla, where large multi-sensor real-world datasets were what made driverless cars work. [3] That post also set out a research goal that Explorer did not reach: a single 3D representation unifying the editability of meshes, the appearance quality of splats and the real-world learning of neural radiance fields. [3] The post is no longer listed on Odyssey's current research index. [9][10]

Explorer is the only Odyssey model described by the company as purpose-built rather than general. Its research index entry still reads "A purpose-built world model for film and gaming". [9]

Odyssey-1

Odyssey published the model now called Odyssey-1 on 28 May 2025 as a free public research preview. [4][6] The claimed capabilities were spatial consistency, actions learned from video, and coherent video streams of five minutes or more, with a new frame generated and streamed every 40 milliseconds. [4]

The serving details were published in unusual specificity. Odyssey said the preview streamed at up to 30 FPS from clusters of H100 GPUs in the United States and the EU, that end-to-end latency from keypress to frame could be as low as 40 ms, and that the infrastructure cost was $1 to $2 per user-hour depending on video quality. [4][6] Oliver Cameron told The Decoder that "every frame is absolutely generated by a diffusion model we've trained". [7]

Odyssey was candid that the model was narrow. To keep long autoregressive rollouts stable it pre-trained on general video and then post-trained on video from a small set of densely covered places, trading generality for stability. [4] The company's own description was that it felt like "exploring a glitchy dream", which it called "raw, unstable, but undeniably new". [4] TechCrunch confirmed the wobble from the outside, reporting blurry and distorted environments whose layouts did not always stay the same when the viewer turned around. [6]

The post positioned the work against game-world research such as Decart's Oasis and Microsoft's WHAMM, arguing that Minecraft-style and Quake-style environments impose a low ceiling because their pixels, motion, actions and physics are all constrained, and that learning from real-life video would lift it. [4] It also announced that a next-generation model with stronger generalisation was already producing outputs. [4]

Odyssey-2 Alpha

Odyssey-2 Alpha is the least documented model in the line. It appears on Odyssey's research index between Odyssey-1 and Odyssey-2, and the description there has changed: the version archived on 12 March 2026 read "A world model with navigation input and generality", while the current index reads "A general world model with navigation input". [10][9] The index entry does not link to an announcement post; it links to a researcher video page. [9] Odyssey's blog index, whose earliest entry is the December 2024 Explorer post, has no entry for it. [21]

Given the descriptions, Odyssey-2 Alpha is most plausibly the "next-generation world model" whose outputs were previewed inside the Odyssey-1 post in May 2025, but Odyssey has not published a date, a specification or an availability statement for it. [4][9]

Odyssey-2

Odyssey-2 was announced on 27 October 2025. [11][12] The step change was the input: where Odyssey-1 took navigation actions, Odyssey-2 took open-ended text typed during a live stream. Odyssey's own framing was that you experience it like a language model, typing while the video responds in the moment. [11]

Odyssey described the training recipe only in outline: the model starts from a bidirectional video model with general knowledge of the world, then goes through "a novel multi-stage training pipeline that transitions the model to causal behavior", producing a real-time, action-aware model that generates each frame from prior frames and user input alone. [11] The argument for why causality matters is the clearest statement of Odyssey's position: a bidirectional video model builds an embedding of a whole clip in advance, so the ending is fixed before the viewer can act on it. [11][15]

Published performance figures for Odyssey-2:

Property	Value
Frame interval	Under 50 ms [11]
Frame rate	20 FPS [11]
Session length	Multi-minute streams [11]
Input	Text typed during the stream [11]
Comparison Odyssey drew	Bidirectional video models taking 1-2 minutes to produce 5 seconds of footage [11]

At launch the API did not exist; the post said "our API is coming soon". [11][12] A free web demo was available. [12]

Odyssey-2 Pro and the Odyssey API

Odyssey-2 Pro shipped on 23 January 2026 alongside the company's first public developer API, in a post titled "The GPT-2 Moment for World Models Is Here". [13] Odyssey described Odyssey-2 Pro as significantly larger than Odyssey-2, with the extra capacity spent on physics, dynamics, behaviour and pixel quality, and said it streamed 720p video at 22 FPS. [13] eWeek, covering the launch independently, reported the same 720p and 22 FPS figures. [14]

The API exposed three endpoints, which Odyssey described as follows. [13]

Endpoint	What it does
Simulations	Takes a prompt, a set of actions at specified time steps, quality settings and a target length, and returns a video
Interactive Streams	Embeds a stream that the host application can interact with programmatically in real time
Viewable Streams	Distributes a single interactive stream to many viewers

JavaScript and Python SDKs shipped at launch, with iOS and Android SDKs described as coming. [13] Odyssey later added an API feature called Broadcast, described as letting multiple users join the same live simulated experience; shared simulation became the subject of a separate model, Agora-1, in May 2026. [22][21]

Odyssey-2 Pro became the company's shop window for the next few months. It was the model named in the 12 February 2026 announcement of investment from NVentures, NVIDIA's venture arm, and Samsung Next, where Odyssey called it "a breakthrough general-purpose world model, which developers can now integrate", and where Samsung Next investment director Andy Duong was quoted praising "the rapid technical advances demonstrated by Odyssey-2 Pro". [19] It was also the model used as the running example in Odyssey's December 2025 essay "The Dawn of a World Simulator". [20]

Odyssey-2 Max

Odyssey-2 Max was announced on 21 April 2026 as the third model in the Odyssey-2 family and the largest the lab had trained. [15] Odyssey put the scaling at roughly 3x the parameter count and 10x the training compute of Odyssey-2 Pro, and said it trained on several hundred NVIDIA Blackwell B200 GPUs. [15]

Odyssey published an architecture summary rather than a paper. The model is an autoregressive diffusion transformer. The design choices Odyssey listed were a proprietary KV cache enabling real-time inference and training on sequences up to 20x longer than prior work with full backpropagation; causal attention combining local and global context; conditioning on latent-space embeddings so that arbitrary action inputs can be handled; inference-aware modelling using roofline estimates from the start; flow matching in a continuous latent space instead of discrete tokenisation; and distillation to a small number of denoising steps. [15] Training ran in three stages: large-scale video pretraining for general visual dynamics, then interaction and task conditioning, then a long-horizon stability phase. [15]

The benchmark table was the first Odyssey published. The company evaluated on the physics sub-score of VBench-2.0, which assesses mechanics, thermotics, materials and multi-view consistency, and on the physics subset of PAI-Bench, attributing baseline numbers to Zheng et al. (2025) and Zhou et al. (2025) alongside its own measurements. [15][17][18] Odyssey's reported figures:

Model	Class	VBench 2 physics	PAI-Bench physics
Odyssey-2 Max	General world model	58.52	93.02
Odyssey-2 Pro	General world model	49.67	91.67
Odyssey-2	General world model	48.58	89.50
Cosmos-Predict2.5-14B	Physical AI world model	44.92	93.50
Cosmos-Predict2-14B	Physical AI world model	39.22	89.20
Cosmos-Predict2.5-2B	Physical AI world model	35.61	91.70
LingBot-World-Fast	Gaming world model	Not reported	92.19

These are Odyssey's own numbers, and the comparison set is Odyssey's own choice. The company stated that it benchmarked publicly available world models and deliberately excluded bidirectional video models on the grounds that they do not meet its definition of a world model, which removes the strongest video generators from the table. [15] On PAI-Bench physics, Odyssey's own table shows Cosmos-Predict2.5-14B ahead of Odyssey-2 Max (93.50 against 93.02), so the state-of-the-art claim rests on the VBench 2 physics column. [15]

Odyssey-2 Max was not released publicly at launch. The company said it was available in private beta to partners working in robotics, gaming, simulation, defence and interactive systems. [15] By September 2026 the developer portal stated that Odyssey-2 Max was "being rolled out to existing API users", with new developers asked to request access rather than self-serve an API key. [24] Odyssey compared the model's state to "a human who has spent years observing and interacting with the world, just before learning to drive", and to GPT-2 just before ChatGPT. [15] On X the company said the model "materially advances the SOTA in physical accuracy". [16]

The retroactive relabelling

Odyssey did not use the phrase "world model" consistently across this period, and the current versions of its older posts do not reflect what they originally said. This is visible in the Internet Archive.

The December 2024 Explorer post was published as "Generative World Models for Film, Gaming, and Beyond" and called Explorer "our first generative world model"; the word "generative" has since been removed from the title and body. [1][25]
The 28 May 2025 launch was published under the title "AI video you can both watch and interact with in real-time", describing "a new interactive video model" and "an entirely new form of media" that Odyssey called interactive video. The archived copies from 27 October 2025 and 21 January 2026 still read that way, at the URL /introducing-interactive-video. By 20 February 2026 that URL returned a redirect and /introducing-odyssey-1 served the retitled version. [5][4]
The 27 October 2025 Odyssey-2 post was published as "Introducing Odyssey-2: instant, interactive AI video" and called the model "a frontier interactive video model". It is now titled "Introducing Odyssey-2: A General-Purpose World Model" and calls it "a frontier world model". Section headings changed with it: "Interactive video models are burgeoning world simulators" became "World Models Are Burgeoning World Simulators". [12][11]
The find-and-replace left a visible seam. The live Odyssey-2 page contains the fragment "unlocked reasoning and creativity. models take that idea further." where the archived launch-day text reads "Interactive video models take that idea further." [11][12]
The research index descriptions were also revised. In March 2026 Odyssey-2 was "A world model with language input and generality" and Odyssey-2 Pro carried the same line; both now read "A general world model with open-ended inputs". [10][9]

None of this changes what the models did. It does mean that any date, name or self-description taken from the current odyssey.systems posts should be checked against an archived copy before being treated as contemporaneous.

Reception

Explorer drew coverage focused on the Pixar connection and on the practical limits. TechCrunch compared it to world models from DeepMind, World Labs and Decart, reported the 10-minute generation time and the artifacts, and noted that Odyssey had raised $27 million to that point from EQT Ventures, GV and Air Street Capital. [2]

Odyssey-1 got the widest coverage of any model in this line. TechCrunch, eWeek and The Decoder all covered the 28 May 2025 preview, and all three described it as a research demo rather than a product. [6][7][8] eWeek called Odyssey "London-based" and led on the Pixar board connection. [8] TechCrunch set the launch against the wider debate about generative AI in entertainment, citing an Animation Guild-commissioned 2024 study estimating that more than 100,000 United States film, television and animation jobs would be disrupted by AI, and noted Odyssey's pledge to collaborate with creative professionals rather than replace them. [6]

Odyssey-2 Pro's API launch was covered by eWeek, which paired it with a World Labs API release and endorsed Odyssey's own GPT-2 framing. [14]

Relation to Odyssey-3

Odyssey-3, announced on 15 September 2026, is the successor to Odyssey-2 Max and is described by the company as a foundation world model aimed at powering robots, driving cars and training other AI systems. [23] Between Odyssey-2 Max and Odyssey-3 the lab shipped work that branched away from the single-stream, single-user Odyssey-2 design: PROWL-1 in May 2026, an adversarial reinforcement learning framework for improving world models; Starchild-1, which added audio; Agora-1, which put multiple participants in one simulation; and CaliBench, a benchmark for whether world models reproduce the true distribution of physical outcomes. [9][21] The Odyssey-2 family remains the lab's bridge from a creative-tools product to a general world model research programme.

References

World Models for Film, Gaming, and Beyond - Oliver Cameron, Odyssey, 18 December 2024.
AI startup Odyssey's new tool can generate photorealistic 3D worlds - Kyle Wiggers, TechCrunch, 18 December 2024.
A new frontier for generative models: learning from our world - Odyssey, 13 November 2024, archived copy.
Introducing Odyssey-1: A Playable World Model - Oliver Cameron, Odyssey, dated 28 May 2025 (current version).
AI video you can both watch and interact with in real-time - Odyssey, 28 May 2025, archived 27 October 2025.
Odyssey's new AI model streams 3D interactive worlds - Kyle Wiggers, TechCrunch, 28 May 2025.
Generative AI startup Odyssey demos interactive AI-generated video - The Decoder, May 2025.
'Undeniably New' AI-Generated Interactive, Real-Time Video From Startup With Pixar Connection - eWeek, May 2025.
Leading World Model Research - Odyssey research index, retrieved 16 September 2026.
Odyssey research index - archived 12 March 2026.
Introducing Odyssey-2: A General-Purpose World Model - Oliver Cameron, Odyssey, dated 27 October 2025 (current version).
Introducing Odyssey-2: instant, interactive AI video - Odyssey, archived 27 October 2025.
The GPT-2 Moment for World Models Is Here - Oliver Cameron, Odyssey, 23 January 2026.
World Models Just Got Their GPT-2 Moment - eWeek, January 2026.
Introducing Odyssey-2 Max: Scaled World Simulation - Oliver Cameron, Odyssey, 21 April 2026.
Odyssey (@odysseyml) on X - 21 April 2026.
VBench-2.0: Advancing Video Generation Benchmark Suite for Intrinsic Faithfulness - Dian Zheng et al., arXiv:2503.21755, March 2025.
PAI-Bench: A Comprehensive Benchmark For Physical AI - Fengzhe Zhou et al., arXiv:2512.01989, 2025.
Odyssey Announces Investment from NVentures and Samsung Next - Oliver Cameron, Odyssey, 12 February 2026.
The Dawn of a World Simulator - Oliver Cameron, Odyssey, 20 December 2025.
The Latest from Odyssey - Odyssey blog index, retrieved 16 September 2026.
Say Hello to Broadcast - Odyssey, archived 20 April 2026.
Introducing Odyssey-3: A General-Purpose Physical Intelligence - Oliver Cameron and Jeff Hawke, Odyssey, 15 September 2026.
Odyssey developer portal - Odyssey, accessed 16 September 2026.
Odyssey homepage, archived 18 December 2024 - Internet Archive Wayback Machine.

Editing: Odyssey-2