Odyssey-3: What Odyssey's Foundation World Model Actually Is

The first thing to get right about Odyssey-3 is the noun. It is not a video generator that has been branded a world model for marketing reasons — Odyssey Systems describes it, on its own announcement page, as "a learned dynamical system, implemented as an autoregressive diffusion transformer, that predicts how objects move and interact through space and how situations evolve over time." The two-tier release, Odyssey-3 at 832×480 and Odyssey-3 Pro at 1280×720, is a model you act inside rather than a model you prompt once and walk away from. Odyssey announced the work on September 15, 2026 under the title "A General-Purpose Physical Intelligence" and opened a public research preview on October 8, 2026; this page is the reference entry for the model itself, not a report on either date.

What that means in practice: on Odyssey's own page you type a prompt, the model builds an environment, and you move through it — first-person, third-person, or as a free-floating camera — while the model keeps predicting what happens next. Odyssey calls the distilled, few-step variant at the centre of that experience "capable of real-time interaction." Nothing about it is priced, licensed, or downloadable. There is no model card, no weight release, and no self-serve key. That combination is the whole of the reader's question, and it is worth taking in the order the decision actually needs it.

What class of model is Odyssey-3?

Readers searching for "Odyssey 3 world model" have the classification right, and it is Odyssey's own term: the page headline calls Odyssey-3 its "most powerful foundation world model yet." The precise class is narrower than "world model" in general use. Odyssey-3 is an interactive, action-conditioned environment model — it generates an embodied environment from a prompt and then predicts how that environment changes as a person or agent takes actions or introduces events, conditioned on previous observations plus the latest input. Odyssey's wording is that today's preview "provides first-person and third-person navigation alongside independent camera movement."

Why that is a different artifact from a video model, in Odyssey's own account of how it was built:

• It is built as a multi-step video diffusion transformer with temporally resolved prompts and controls, then extended autoregressively through teacher forcing and causal masking so it can continue from preceding observations and predict future states conditioned on action inputs.

• A post-training pipeline combining distribution-matching and adversarial distillation produces a few-step distilled variant — the piece that makes real-time interaction possible at all, since a diffusion model that needs dozens of denoising steps per frame cannot keep up with a keypress.

• It is trained on internet video carrying time-localized, schema-verified event annotations, on gameplay recordings with time-aligned keyboard and mouse inputs, and on simulated rigid-body interactions with captions and metadata. The gameplay and simulation sources are what attach actions to observations; the internet video supplies the breadth.

• Output is video, and the input surface includes text prompts, uploaded images, and continuous movement or camera control. It is not a language model and has no context window in the LLM sense; nothing in Odyssey's material describes one.

So the correct one-line answer is: a world model, specifically an interactive environment model that outputs video — not a video generator, and not a chatbot-adjacent multimodal LLM. The twelve comparison pages on this site that set it against Seedance, Veo, Kling, Runway and similar are comparing an environment against clip generators; the benchmarks below are where that distinction actually shows up in numbers.

What you can do with it today

The research preview is live at experience.odyssey.systems, and Odyssey's page states plainly that "our research preview is available now." Reading the preview application itself, the controls are:

• A prompt box ("Describe what you'd like AI to simulate"), an image upload, and a gallery of suggested scenes that you browse with the arrow keys and pick with Enter.

• Three starting camera intents — "World View," described in the interface as "a free-floating camera with no player attached"; "First-Person View," "see through the playable subject's eyes"; and "Third-Person View."

• A start control that begins simulation, after which you can move through the environment and introduce events mid-generation and watch how the model responds.

Two things about access you should not have to discover the hard way. First, the button in the preview reads "Try Odyssey-3 Flash" — a tier name that appears nowhere else on the site. Odyssey names exactly two tiers, Odyssey-3 and Odyssey-3 Pro, and publishes no specifications, size, or resolution for anything called Flash. Treat the label as unexplained until Odyssey documents it. Second, the API path is not self-serve. The "Get API Access" link points at developer.odyssey.ml, and that page currently reads: "Odyssey-2 Max is now being rolled out to existing API users. If you already have an account with us, please continue. Otherwise, head here to request priority access." The developer surface being advertised today is Odyssey-2 Max, not Odyssey-3. The route to Odyssey-3 for a builder is the contact form on Odyssey's own site, and the honest description of it is a queue.

The limits Odyssey documents — and the ones it does not

This is where a reference page earns its keep, because the gaps are as decision-relevant as the figures.

• Resolution — Odyssey-3 832×480; Odyssey-3 Pro 1280×720. Vendor-published, stated directly on the announcement page.

• Frame rate — not published by Odyssey. The only frame-rate figure anywhere in the primary material is the Physics-IQ Verified leaderboard's "Output FPS" column, which lists 16 for every Odyssey entry against 24 for FLUX 3 [large]. That is a third-party field on a third-party board, not a vendor specification, and it is the number the board uses to normalise cost.

• Clip duration — not published. Odyssey documents no maximum generation length, no clip length, and no session length. WorldMark, the benchmark Odyssey cites for world-model quality, runs its own test cases at 20, 40 and 60 seconds — those are the benchmark's probe lengths, not a documented Odyssey-3 limit, and conflating them would be a mistake.

• Interactivity limits — not published as limits. Odyssey documents the modes available (first-person, third-person, free camera, mid-generation events) but publishes no latency figure, no maximum concurrent actions, and no statement of how long a session can run before the world degrades.

• Cost model — the only financial figure Odyssey publishes is an assumption, not a price: it "assumes $1 per MI355X GPU-hour, excluding prompt-rewriting fees" when it computed the cost axis of its Physics-IQ chart. There is no price list, no per-second rate, and no credit scheme on any Odyssey page as of today.

Weights, code and licence: none of the three

Straight answer, on all three counts. No weights have been released — a search of Hugging Face for Odyssey Systems' models returns nothing from the company, and no Odyssey page mentions a checkpoint, a repository, or a download. No training or inference code has been published. And no licence has been published for Odyssey-3; the only licence-and-terms-shaped document on the site is the API Terms page, last updated 2026-01-22, which still governs "the prototype of Odyssey-2 application programming interface" — a document that predates Odyssey-3 by two model generations and does not name it.

If a page anywhere tells you Odyssey-3 is open-weights, or quotes a licence for it, that page is guessing. Nothing in Odyssey's own material supports either claim today.

What Odyssey claims, and what has actually been measured

Odyssey makes one headline benchmark claim: "Odyssey-3 Pro sets a new state of the art on Physics-IQ Verified's video-to-video benchmark, achieving 66.1, the highest reported score," plus a second claim that Odyssey-3 "ranks 1st in 3 of WorldMark's 4 categories in our evaluations." Those two claims have different warrants and should be read apart.

Physics-IQ Verified is a genuine third-party board. It is run by Anates Labs with DeepMind, it currently lists 41 models from 16 labs, and it scores models by showing them video of real physical experiments — fluid dynamics, optics, solid mechanics, magnetism, thermodynamics — and measuring their continuation against what actually happened. Reading the video-to-video verified board on 2026-10-13:

• Odyssey-3 Pro, BoN ×8 sampling — 66.10, rank 01. Odyssey-3, BoN ×8 — 64.43, rank 02. FLUX 3 [large], BoN ×8 — 64.35 ± 0.22, rank 03. The vendor's 66.1 figure matches the top row on the board, so the "highest reported score" claim holds up on the board's own current ranking.

• Single-generation rows are lower and closer together: Odyssey-3 Pro 63.37 ± 0.63, Odyssey-3 61.56 ± 1.20, FLUX 3 [large] 61.11 ± 0.55. The ordering is not stable across sampling. Under best-of-8 the base model's 64.43 sits 0.08 points clear of FLUX 3's 64.35 and Pro's 66.10 sits 1.75 clear. Under single generation FLUX 3 lands between the two Odyssey rows — 0.45 below Odyssey-3's 61.56 and 2.26 below Pro's 63.37.

• The most important row for a sceptical reader is further down. Odyssey-3 also appears with the prompt source marked "BPP" — benchmark-provided prompts, no LLM rewriting — at 51.76 ± 0.69, rank 08. Its custom-prompt row is 61.56. On the benchmark's own prompts, the model's verified score is roughly ten points lower than on its own. The board is explicit that separate LLM/prompt costs are added after generation-cost normalisation, and that Odyssey API prompt rewriting is estimated at $0.01 per submitted video, with each best-of-8 run sharing one rewritten prompt.

• On the image-to-video verified board the ranking is different and the vendor's framing should be adjusted accordingly: FLUX 3 [large] leads at 54.70 (BoN ×8), with Odyssey-3 Pro second at 54.69 and Odyssey-3 third at 52.83. A tenth of a point is not a lead in either direction. Odyssey's claim of state of the art is a video-to-video claim, and on image-to-video it is not the leader.

• Normalised cost, which the board computes to 24 fps and 1280-wide output where the inputs are reported: on the video-to-video best-of-8 rows Odyssey-3 Pro is $3.484 and Odyssey-3 is $1.430 against FLUX 3 [large] at $16.430; on the single-generation rows, Odyssey-3 Pro is $0.444 and Odyssey-3 is $0.187. Note that earlier coverage of this launch quoted $0.139 and $0.267 for the two Odyssey tiers — those are the image-to-video single-generation rows on the same board, not the video-to-video rows, and the two should not be swapped. The figure did not move; it belongs to a different sampling row.

• Submetrics, video-to-video verified: Odyssey-3 Pro leads spatial at 73.19 and weighted spatial at 65.07; Odyssey-3 leads spatiotemporal at 70.44. FLUX 3 [large] leads mean squared error at 57.39, with Odyssey-3 Pro second at 55.81. Odyssey is ahead on the spatial side and not on error.

The WorldMark claim has a weaker warrant, and this is the part to read carefully. WorldMark is a real, public benchmark suite from Alaya Lab, the University of Tokyo and Shanghai Innovation Institute: it drives ten models from one shared WASD-style action vocabulary over 500 standardised cases and scores nine deterministic metrics across action dynamics, world memory and visual quality. Odyssey's framing is that "in our evaluation, using the benchmark's own captions and the mean of its 13 reported metric scores, Odyssey-3 ranks first in first-person stylized, third-person real, and third-person stylized environments." Three things follow from the primary material:

• It is a self-run evaluation. Odyssey ran the benchmark and reported the result; WorldMark's authors did not publish it.

• Odyssey-3 does not appear in the benchmark's own published results. Reading WorldMark's results page on 2026-10-13, the four splits list ten models — AlayaWorld, DreamX-World, HY-GameCraft 1.0, HY-World 1.5, LingBot-World, Lyra 2.0, Matrix-Game 2.0, Matrix-Game 3.0, SANA-WM and Yume 1.5 — and Odyssey-3 is not among them. So the ranking cannot be checked against the benchmark's own numbers, because on the public board there are none.

• The metric count Odyssey describes differs from the benchmark's own description. WorldMark's page describes "nine deterministic metrics" in three categories; Odyssey describes "the mean of its 13 reported metric scores." Both can be true of different views of the suite, but a reader cannot reconcile them from the outside, and the number that changes is the top-line one.

• The one split Odyssey-3 does not win in its own evaluation is the one closest to the product pitch: first-person real environments, where Odyssey's figure of 80.6 trails Lyra 2.0 at 84.4 and AlayaWorld at 83.0. The model's stated edge is concentrated in stylised and third-person worlds.

Beyond the two benchmarks, every capability claim is vendor-reported and independently unreproduced, and the reference page should say so without softening it. Odyssey reports a robot-arm action decoder trained on "only tens of hours of robot demonstrations" that produced recovery behaviours absent from the demonstrations, including reorienting a gripper after a missed grasp. It reports humanoid control policies built by Flexion on Odyssey-3 that outperformed the tested VLA baselines under environmental change, and continued under lighting changes that broke those baselines. It reports a driving policy trained on 20 hours of driving data with the Odyssey-3 backbone held frozen, driving in closed loop on real roads in India — and the single comparative figure in that work is that simulation-only policies travelled about 77% as far between safety-driver interventions as policies trained on real footage. It reports multi-sensor generation, with an Odyssey-3 checkpoint producing the front three camera views of a driving dataset together after 100 training steps. None of it has been rerun by anyone outside Odyssey, and no third party has published a number for Odyssey-3 that Odyssey did not supply.

There is also, as yet, no head-to-head to lean on. Nobody has run Odyssey-3 against FLUX 3 [large] or any other model at matched harness and matched sampling outside the Physics-IQ submission itself, and the Physics-IQ table is precisely where the sampling choice — best-of-8 versus single generation, custom prompts versus benchmark prompts — moves the result by ten points. Treat any ranking that does not name its sampling row as uninformative.

How to get it, and what to do meanwhile

Today the routes are the browser preview at experience.odyssey.systems and the contact form on odyssey.systems for API access. That is the complete list; there is no pricing page, no rate-limit documentation, and no self-serve key. Odyssey-3 is not something we can route for you either — there is no callable API to point at yet, so any availability line aimed at developers is premature.

If that leaves you needing something you can call this week, the useful move is architectural rather than editorial: keep the model behind an interface you control, so that dropping Odyssey-3 in later is a configuration change rather than a rewrite. That is the case routing layers exist for, and it is also the correct way to evaluate a model whose benchmark numbers move ten points depending on sampling — you want to be able to A/B the real thing against your incumbent on your own prompts, on one key, without a second contract. OrcaRouter puts more than 200 models behind a single endpoint, passes provider list price through at 0% markup, and fails over automatically across providers — which matters most for exactly the situation Odyssey-3 is in: a model you cannot yet get, next to models you can.

What to watch, concretely, and in order of how much it would change the answer: a published price and a self-serve key; a licence and a weight release, or an explicit statement that neither is coming; a frame-rate and duration specification from Odyssey rather than from a leaderboard column; and Odyssey-3 appearing in WorldMark's own published results, which would turn a self-run ranking into a checkable one. Until at least the first and last of those land, the accurate description of Odyssey-3 is a genuinely interesting foundation world model — an interactive environment model that outputs video, not a video generator — with one strong third-party benchmark result under best-of-8 sampling, a much more modest one under the benchmark's own prompts, and no commercial surface at all.