Odyssey interactive video (world models)

What this skill is for

Odyssey builds world models: systems that generate video one frame at a time in real time, conditioned on the history of frames plus whatever input a user is streaming right now. The output is not a finished clip you download — it is an interactive stream the user drives. This skill helps an agent scope, plan, and honestly advise on projects that use (or consider using) Odyssey's models.

Use it when a request looks like:

• "Can we build an explorable AI world / playable dream the user walks through?"
• "Real-time interactive video that reacts to typed prompts or a controller."
• "Should we use Odyssey or Sora/Veo for this?" (interactive vs. offline decision).
• "Prototype an interactive environment / experiential installation / live demo."

Do not route here for: offline text-to-video clips you render once and download (that is Sora/Veo/Kling/Runway territory), exportable 3D assets or meshes (that is a 3D-asset or game-engine pipeline), or captioning/editing existing footage. Odyssey produces no downloadable clip and no exportable geometry by default — if the deliverable is a file to hand off, interactive video is usually the wrong tool. See "Production decision" below.

Maturity flag, read first. Odyssey's own CTO Jeff Hawke describes the field as "the GPT-2 era of world models" and "a phase of mass exploration, not mass commercialization" [Secondary, The Data Exchange, verified 2026-07-10]. Treat every capability here as research-grade. Prototypes, demos, art, and experiments are feasible today; SLA-bound production pipelines generally are not. Say this plainly to any user with a production deadline.

Evidence labels used below

• [Documented] — stated in Odyssey's own model pages / technical reports.
• [First-party claim] — Odyssey's own performance/capability assertion (marketing or blog), not independently verified.
• [Secondary] — reputable press or analyst reporting.
• [Heuristic] — production judgment from the nature of the technology, not a cited fact.

All volatile facts (model names, latency, access, pricing) are dated verified 2026-07-10 unless noted. Re-verify before relying on them — this product changes on a scale of weeks.

What "interactive video" means operationally

A traditional video model (Sora, Veo, Kling) is bidirectional: it plans a whole clip with knowledge of its own future frames, so it takes seconds-to-minutes to produce a few seconds of fixed video. You cannot change it mid-generation.

Odyssey's models are causal and autoregressive: each new frame is predicted only from past frames and past/incoming user actions, then streamed immediately [Documented, odyssey.ml/introducing-odyssey-2]. Consequences an agent must internalize:

1. The user is in the loop continuously. Input (keypress, mouse, typed text, speech, controller) is fed in mid-stream and changes what comes next. Odyssey-1 explicitly likens this to "reacting to what happens"; Odyssey-2 likens prompting during playback to "conversing with a language model" [Documented].
2. Latency is per-frame, not per-clip. Odyssey-1: a new frame roughly every 40 ms (up to ~30 fps) [Documented, 2025-05-28]. Odyssey-2: ~50 ms/frame, ~20 fps [Documented]. Starchild-1: synchronized audio+video up to 24 fps [Documented, 2026].
3. Every frame is generated, not composited. Odyssey states "every frame is absolutely generated by a diffusion model we've trained" — no game engine, no pre-authored 3D assets, no meshes [Documented/Secondary, TechCrunch 2025-05-28].
4. Coherence is bounded and drifts. Streams stay plausible for a bounded window, then degrade. See "Capability limits."

Current model lineup (verified 2026-07-10)

Odyssey (odyssey.ml) is a world-model research lab; founders Oliver Cameron (CEO) and Jeff Hawke (CTO) come from autonomous driving, and Pixar co-founder Ed Catmull sits on the board [Secondary, TechCrunch]. The public model family, oldest to newest:

Model	First public	What it is	Interactive input	Frame rate / latency	Access route
Odyssey-1	2025-05-28	First "playable world model"; action-conditioned next-frame predictor. Narrow-distribution, deliberately traded generality for autoregressive stability. Described by Odyssey as "a glitchy dream — raw, unstable."	keyboard / touch / joystick	~40 ms/frame, up to 30 fps, 5+ min coherence claimed	Superseded (research preview)
Odyssey-2 (API tier: Odyssey-2 Pro)	2025-10-27 blog	"General-purpose world model that dreams AI video instantly you can interact with." Steered by natural-language prompts during playback.	typed natural-language prompts (+ actions)	~50 ms/frame, ~20 fps	Web: experience.odyssey.ml · API: developer.odyssey.ml (emerging)
Starchild-1	2026-05-17	First multimodal world model — synchronized audio + video, generated autoregressively; environments, dialogue, ambient sound evolve from streaming input.	streaming text, speech, or action	up to 24 fps, synced A/V	Preview + technical report at odyssey.ml (starchild.odyssey.ml/starchild-1.pdf)
Agora-1	2026	Multi-agent world model: up to 4 players in one shared simulated environment (a GoldenEye-style deathmatch demo), each player's viewpoint rendered live. Decouples world-state simulation from per-view rendering.	multiple players' actions	real-time per-view	Playable at agora.odyssey.ml
PROWL-1	2026	Reinforcement-learning framework for exploring game environments (research paper, arXiv 2605.18803). Research artifact, not an interactive product.	n/a	n/a	arXiv paper

[All rows Documented from odyssey.ml model pages / homepage, verified 2026-07-10.]

Naming heuristic for agents: "Odyssey" alone in 2026 usually means Odyssey-2 (the live general-purpose experience). If a user says "the interactive video thing from Odyssey," default to Odyssey-2 and confirm. Starchild-1 is the one to name when sound matters; Agora-1 when multiplayer/shared state matters.

Access routes today — verify carefully, do not assume an API

This is the area most likely to be misstated. As of 2026-07-10:

• Web experiences are the reliably public route. experience.odyssey.ml (Odyssey-2) and agora.odyssey.ml (Agora-1) are publicly playable in a browser [Documented, odyssey.ml homepage]. This is the safest thing to point a user to today and the right way to evaluate the model before committing.
• An API exists but is early/emerging. There is developer-facing documentation at documentation.api.odyssey.ml and a dashboard at developer.odyssey.ml/dashboard describing an Odyssey-2 Pro API with three surfaces: interactive streams (real-time), viewable/broadcast streams, and a Simulate endpoint for offline/batch scripted generation [Documented, API docs, verified 2026-07-10]. Authentication is via an API key (keep server-side; browsers connect with short-lived credentials). It references session/concurrency limits and stream-duration limits (a "credit lease expired" error), a Discord, an API License Agreement, and support@odyssey.ml.
  • Caveat an agent must surface: Odyssey's own homepage did not prominently advertise the API on 2026-07-10, and the Odyssey-2 launch blog originally said the API was "coming soon." So treat API availability as real but nascent — confirm current status, quotas, and terms directly at developer.odyssey.ml before promising an integration. Do not assume stable endpoints, versioned SDKs, or uptime guarantees.
• Starchild-1 and PROWL-1 are, as of this date, research/preview + technical report — a report and preview, not a documented general API tier. Do not promise programmatic Starchild-1 access without checking.

[Heuristic] What to tell a user about access: "You can try it yourself right now in a browser at experience.odyssey.ml. Programmatic/API access via Odyssey-2 Pro is available but early — I'd verify current quotas and terms at developer.odyssey.ml before we design an integration around it."

Capability limits (the honest section)

Every one of these is load-bearing for scoping. [Documented/First-party unless tagged.]

• Visual fidelity is unstable. Independent and Odyssey descriptions alike report blurred, distorted, dreamlike output; fine detail and text-in-scene are unreliable [Secondary, TechCrunch/eWeek; First-party "glitchy dream"].
• Coherence is time-bounded and drifts. Odyssey claims "5 minutes or more" of coherent stream for Odyssey-1; Hawke's later framing is more conservative — "one to two minutes" before degradation, up from an earlier "15 to 30 seconds" [First-party / Secondary]. [Heuristic] Plan around ~1–2 minutes of reliable coherence, not 5+, for anything you show a client.
• World persistence is weak. Because state lives in the model's rolling context rather than a stored scene graph, geometry and layout can shift when you look away and back; objects are not guaranteed to persist. Starchild-1's report explicitly notes "scene and acoustic identity still drift over long horizons" [Documented]. There is no reliable "save this world and return to the exact same place" primitive.
• Resolution / frame rate are modest. ~20–30 fps; resolution is not headlined by Odyssey and is well below film-finishing quality [Documented for fps; resolution not published — do not quote a number].
• No exportable asset. You get streamed frames (and, for Starchild-1, audio), not a mesh, not a texture, not a downloadable MP4 clip by default. The Simulate endpoint produces recordings, but the artifact is generated video, not editable 3D [Documented].
• Per-user GPU cost and infra. Runs on GPU clusters (H100s historically; AWS/Trainium going forward). Odyssey's own 2025 figure was $1–$2 per user-hour of compute, expected to fall [First-party, 2025-05-28]. Every concurrent viewer consumes a live GPU session — this does not scale like serving a cached video file.
• No quantitative benchmarks for the interactive style yet. Starchild-1's report says standardized benchmarks for this interaction mode "don't yet exist" [Documented]. You cannot cite an FVD/quality number to a stakeholder.

Production decision: interactive video vs. the alternatives

Choose the tool by what the deliverable actually is. [Heuristic framework.]

Pick Odyssey interactive video when:

• The core experience is live, open-ended user steering — the user's real-time input must change the video, and no fixed script would do.
• You are building a demo, prototype, art piece, research probe, or experiential/live installation, and dreamlike/unstable aesthetics are acceptable or even desired.
• Novelty and "the user drove this themselves" is the point (marketing activation, conference booth, playable teaser).

Pick an offline video generator (Sora / Veo / Kling / Runway) instead when:

• You need a finished, downloadable clip with controlled composition, fixed length, and repeatable output.
• Visual fidelity, brand-safe detail, legible text, or client-approvable frames matter.
• Distribution is one-to-many (a file served to thousands) — offline clips scale cheaply; interactive streams do not.

Pick a game engine / downloadable 3D (Unreal, Unity, Gaussian splats, NeRF-to-mesh) instead when:

• You need persistent, deterministic worlds, exact collision, exportable assets, or the same environment every session.
• You must ship an asset that runs on the client without a live GPU backend per user.
• Long sessions (hours) or guaranteed geometry are required.

[Heuristic] One-line litmus test: Does the user need to steer the pixels live, and can you tolerate a dreamlike, non-persistent, GPU-per-viewer stream? Yes → Odyssey. Any "no" → probably an offline or engine pipeline.

Feasible now vs. speculative

Feasible today (research-grade, expectations set):

• Interactive prototypes and concept demos — "walk through this vibe" explorable moodboards.
• Experiential marketing / live activations where the novelty of real-time AI worlds is the draw and a staffed, time-boxed session hides drift.
• Interactive storytelling / "playable dream" art that leans into the surreal, unstable aesthetic rather than fighting it.
• Rapid ideation — quickly feeling out an environment or mood before committing to a real 3D/production build.
• Odyssey's own cited proof points: hackathon-style interactive weather apps, game scaffolding [Secondary, Hawke].

Speculative / not reliable yet (do not promise):

• Shippable AAA-style games or engine replacements.
• Brand-safe, pixel-controlled advertising deliverables requiring legible logos/text and approvable frames.
• Persistent multiplayer worlds with durable state (Agora-1 is an early multi-agent demo, not a live-service platform).
• Long-form (hours) coherent experiences.
• Anything needing an exportable asset, SLA, or guaranteed reproduction of the same world.

Pricing and access terms

• No published consumer price tiers as of 2026-07-10. Pricing signal is usage-based and tied to live GPU consumption; there is a credit/lease model referenced in the API docs ("credit lease expired") [Documented].
• Historical compute reference: $1–$2 per user-hour (Odyssey's own 2025 figure, expected to decline) [First-party]. Treat this as an order-of-magnitude cost driver, not a current quote.
• Web experiences are free to try in-browser [Documented].
• [Heuristic] Always get a current quote from developer.odyssey.ml before budgeting — and budget per concurrent-user-hour, because that is the real cost shape.
• Company context (affects longevity risk, not per-project pricing): Odyssey raised a $310M Series B on 2026-06-17 (~$1.45B post-money) led by Natural Capital with Amazon, GV, AMD Ventures, IQT, EQT; AWS is the preferred cloud with Trainium chips [Secondary, multiple outlets]. Well-funded, but still pre-commercialization by its own account.

Content, rights, consent, privacy

• Training data provenance. Odyssey trains partly on its own real-world capture from a proprietary ~25 lb 360° backpack rig (6 cameras, 2 lidar, IMU, ~3.5K, depth metadata), plus other data [Secondary, TechCrunch]. This first-party capture reduces some third-party-scraping concerns but does not eliminate IP questions.
• Output IP is unsettled. Press coverage flags open questions about who owns procedurally generated worlds and whether output resembling known IP could infringe [Secondary]. [Heuristic] Do not assume commercial-use rights, ownership, or indemnity for generated streams — read Odyssey's current API License Agreement / ToS and get commercial terms in writing before client delivery.
• Consent/likeness. If a user streams speech into Starchild-1 or the world depicts real, recognizable people/places, apply normal consent and likeness diligence — real-time generation does not exempt a project from these.
• Privacy of streamed input. Interactive sessions send user keystrokes/speech/actions to Odyssey's servers. For any project handling personal or sensitive input, confirm data-handling/retention terms before deploying.
• Provenance/labeling. Output is synthetic; label AI-generated experiences per platform and jurisdiction rules.

How an agent should plan an Odyssey project

1. Confirm the deliverable is genuinely interactive. If the user actually wants a file to download, redirect to an offline generator or 3D pipeline before going further.
2. Set maturity expectations up front — GPT-2 era, dreamlike, ~1–2 min coherent, non-persistent, GPU-per-viewer. Put this in writing.
3. Pick the model: Odyssey-2 (general, prompt-steered) → default; Starchild-1 → sound needed; Agora-1 → shared/multiplayer.
4. Evaluate hands-on first at experience.odyssey.ml / agora.odyssey.ml before designing anything. Do not design around unseen capability.
5. Verify access route and terms at developer.odyssey.ml (API status, quotas, session/duration limits, license) — dated, because it changes.
6. Budget by concurrent user-hours, not by clip.
7. Design for drift: short sessions, staffed/guided experiences, aesthetics that embrace instability, fallbacks if a stream degrades.
8. Clear rights/consent/privacy before any commercial or public deployment.

Examples (labeled — these are illustrations, not required formulas)

Example A — Scoping request, agent gives an honest routing answer

Intent: A brand agency asks: "We want a 30-second interactive AI world at our booth where visitors type a mood and walk around it. Use Odyssey. Also we want the final clips for social." Good agent response (essential characteristics):

• Splits the two asks. The booth experience is a strong fit for Odyssey-2 at experience.odyssey.ml or via the Odyssey-2 Pro API — real-time, prompt-steered, novelty is the point.
• Flags that "final clips for social" is a different deliverable: Odyssey doesn't hand you a clean downloadable, output is dreamlike/unstable, and it won't be brand-pixel-controlled. Recommend either screen-recording the live session (accepting the aesthetic) or generating the social cut separately with an offline model (Veo/Sora/Kling).
• Sets expectations: ~1–2 min coherent, non-persistent geometry, ~20 fps, GPU cost per concurrent visitor, staffed booth to manage sessions.
• Notes access/terms verification and rights/consent before public use. Why structured this way: the failure mode here is silently promising a clean downloadable social asset from an interactive model. The value is separating "live steer" from "finished file."

Example B — Model selection under a constraint

Intent: "We want an interactive AI experience where a companion character talks back with real voice and ambient sound as the user speaks to it." Expected approach: Sound is central and input is speech, so route to Starchild-1 (synchronized audio+video, streaming speech/text/action input, "narrator/companionship" interaction regime) rather than Odyssey-2 (video-only). Warn that Starchild-1 is preview + technical report stage (verify programmatic access), that acoustic/scene identity drifts over long horizons, and that no interactive-quality benchmarks exist yet. Add consent/privacy handling for streamed speech. Why: picking Odyssey-2 here would silently drop the audio requirement.

Example C — Reality-check on a speculative ask

Intent: "Can we replace our Unity game backend with an Odyssey world model so we don't have to build levels?" Expected approach: No — not for a shippable game today. Explain: no persistent/deterministic world (geometry drifts, no durable state), no exportable assets, no exact collision, coherence bounded to ~minutes, and a live GPU session per player that doesn't scale like a client-side engine. Position Odyssey/Agora-1 as a research probe or prototype for what learned worlds might feel like, and cite Hawke's own "GPT-2 era, not mass commercialization" framing. Recommend a traditional engine for the shippable product. Why: the critical failure is validating an engine-replacement claim the technology cannot support yet.

Sources (verified 2026-07-10)

First-party:

• Odyssey homepage & model index — https://odyssey.ml/ (models: Odyssey-2, Starchild-1, Agora-1, PROWL-1; web experiences).
• "Introducing Odyssey-1: A Playable World Model" — https://odyssey.ml/introducing-odyssey-1 (2025-05-28; 40 ms/frame, ≤30 fps, 5+ min claim, $1–2/user-hour, H100 US/EU).
• "Introducing Odyssey-2: A General-Purpose World Model" — https://odyssey.ml/introducing-odyssey-2 (blog byline 2025-10-27; ~50 ms/frame, ~20 fps, prompt-steered, API "coming soon").
• "Starchild-1: The First Real-Time Multimodal World Model" — https://odyssey.ml/introducing-starchild-1 and technical report starchild.odyssey.ml/starchild-1.pdf (announced 2026-05-17; synced A/V ≤24 fps; long-horizon drift noted).
• Odyssey API documentation — https://documentation.api.odyssey.ml/ and developer.odyssey.ml/dashboard (Odyssey-2 Pro; interactive streams / viewable-broadcast streams / Simulate; API key auth; session & duration limits).

Secondary (labeled as reporting):

• TechCrunch, "Odyssey's new AI model streams 3D interactive worlds" — https://techcrunch.com/2025/05/28/odysseys-new-ai-model-streams-3d-interactive-worlds/ (founders, board, $27M early funding, no game engine, blurry/unstable, $1–2/user-hour, applications).
• TechCrunch, "This generative AI startup is strapping cameras to people's backs" — https://techcrunch.com/2024/11/13/ (360° backpack capture rig specs).
• The Decoder, "Amazon, Nvidia, and AMD bet $310 million…" — https://the-decoder.com/amazon-nvidia-and-amd-bet-310-million-on-ai-startup-building-3d-world-models/ (Series B, AWS/Trainium).
• Air Street Press, "From clip-makers to simulators: Odyssey's new world models" — https://press.airstreet.com/p/odyssey-starchild-1-agora-1 (Starchild-1/Agora-1 detail, interactive vs. offline distinction).
• The Data Exchange, Jeff Hawke interview — https://thedataexchange.media/jeff-hawke-odyssey/ ("GPT-2 era," "1–2 minutes" coherence, mass-exploration-not-commercialization).

Note: model names, latency, fps, access routes, and pricing are volatile; re-verify at odyssey.ml and developer.odyssey.ml before relying on any of them.