Odyssey-3 vs Kling 3: One You Can Call Today, One You Cannot

Odyssey-3 and Kling 3 are easy to line up on a page and hard to compare in practice, because only one of them exists as something you can call. Kling 3 is Kuaishou's video generation model, released in February 2026, generally available through the vendor's API and the Kling app, and — relevant to readers of this blog — routable through OrcaRouter as both kling-v3 and kling-3-turbo. Odyssey-3 is Odyssey's foundation world model, previewed on September 15, 2026, with no published price, no API, and a stated release window of "the coming weeks." So the first honest sentence about this matchup is that it is not a matchup yet: one is a metered product with per-second pricing you can test this afternoon, and the other is a research preview with a demonstrated capability and no way to buy it. The useful question is what each is for, and whether the things Kling 3 does well are things a world model would ever replace.

Kling 3 is a renderer with a storyboarding trick

Kling 3's release in February 2026 put it at the front of the Chinese video-generation pack, and its positioning is unusually concrete. It generates three to fifteen seconds per clip — five seconds by default — at 720p, 1080p, or 4K depending on tier, from text, a single image, or a first-and-last frame pair. What separates it from most of the field is the Omni variant's multi-shot mode: a multi_prompt list of timed shots, each with its own prompt, lets you plan a sequence of cuts in one generation rather than stitching separately generated clips and hoping the lighting matches. Native audio — synchronized speech, ambience, effects, with regional accent support across several languages — is generated in the same pass and can be switched off.

Its documented strengths are human motion and motion control, image-to-video conditioning, and rendering legible on-screen text: signage, logos, product labels that survive the generation. Its documented weaknesses are equally specific — hands and fingers, physical and multi-object simulation, multi-character consistency across a clip, and generation times that run from roughly two to fifteen minutes depending on length and complexity.

That last row of weaknesses is worth pausing on, because it is precisely the row Odyssey-3 is aimed at. Kling 3 struggles with multi-object physical interaction; Odyssey-3's entire thesis is that predicting the next state of a scene forces a model to learn how objects interact. They are not competing claims about the same feature. Kling 3 is being asked to fake physics convincingly enough that a viewer does not notice; Odyssey-3 is being asked to get physics right so a controller can rely on it.

The price you can actually measure

Kling 3's official rates are published per second and start low: $0.084 per second at 720p, $0.112 at 1080p, and $0.42 at 4K, with native audio adding about 50% at 720p and 1080p ($0.126 and $0.168 respectively) and costing nothing extra at 4K. Those figures are the vendor's, and they are the entry-tier rates. Independent provider pricing tracked per minute of output lands well above what those per-second numbers imply, so the honest reading is that Kling 3 is inexpensive to start with and considerably less so once you move up the resolution and audio tiers — check the per-minute figure for your own configuration rather than the headline rate. The consumer side runs on a credit system with tiers from a free daily allowance through to a $127.99-and-up Ultra plan, with the free tier limited to watermarked 720p and no commercial use.

Odyssey-3 has no price at all. That is not a small omission — it is the whole difference in this comparison. A team can compute the cost of a hundred Kling 3 clips before generating one of them. A team cannot compute the cost of anything on Odyssey-3, because there is nothing to compute it from. Note also that if these two ever do appear side by side commercially, they will probably not be billed the same way: video generation meters seconds of output, while world models of the kind Odyssey describes tend to meter hours of simulation, because the buyer is a training loop rather than a render queue.

Odyssey-3 is chasing a different column

Odyssey-3 is an autoregressive diffusion transformer trained on visual observation, and the capability at the centre of the announcement is a learned action decoder — a component that translates the model's internal representation of a scene into the motor commands specific hardware needs. Odyssey reports the same backbone driving robot arms from tens of hours of demonstrations (including unprompted recovery moves like reorienting a gripper after a missed grasp), humanoid policies built with Flexion running in real time, closed-loop driving from 20 hours of simulated data, indoor obstacle-avoiding drone flight, and Grand Theft Auto V sessions that transfer to Red Dead Redemption 2 and Sleeping Dogs without extra policy training in those titles. The company's one comparative figure is that simulation-trained driving policies traveled about 77% as far between safety-driver interventions as policies trained on real footage.

Every one of those claims is unreproduced, there is no Odyssey-3 technical report, and no third party has published a number. The spread across six embodiments is the interesting claim; the absence of a benchmark table is the reason to hold it loosely.

The contrast, dimension by dimension

• What it produces — Kling 3: a rendered clip, up to 15s, at up to 4K. Odyssey-3: a simulated world state plus motor actions.

• Shot structure — Kling 3: multi-shot storyboarding through the Omni variant's timed prompt list. Odyssey-3: a single continuous rollout with no cut concept.

• Audio — Kling 3: native, synchronized, multi-language, optional, priced at +50% below 4K. Odyssey-3: none; it is not a media generator.

• Price — Kling 3: $0.084/s at 720p, $0.112/s at 1080p, $0.42/s at 4K (vendor; audio extra below 4K). Odyssey-3: unpublished.

• Access — Kling 3: GA since February 2026 via the vendor's API and app, and routed on OrcaRouter. Odyssey-3: preview only, "in the coming weeks."

• Evidence — Kling 3: twelve months of public use and third-party pricing documentation. Odyssey-3: vendor demonstrations only.

Where OrcaRouter actually fits — and it is not a footnote here

For once this is not a section about the surrounding stack. Kling 3 is one of the models OrcaRouter serves, as kling-v3 and kling-3-turbo, so the comparison between the two Kling tiers happens on one key rather than two vendor contracts, and you can test both without writing a second integration. Pricing passes through at provider list price with 0% markup, which matters more than usual on a model with a published per-second rate and an audio surcharge: when the vendor moves a number, your bill moves with it the same day instead of at renewal. Automatic failover across providers means a generation queue does not stall because one endpoint is degraded — and video generation endpoints degrade often enough that this is a scheduling concern, not a theoretical one. For pipelines that need several models, the routing DSL composes a prompt model, an image model for the start frame, and the video model into a single call, which is roughly the shape of any real video workflow.

Odyssey-3, by contrast, is not served by OrcaRouter or by anyone else, and this article does not claim it is. There is nothing to route to yet.

The verdict

If you need video this quarter, Kling 3 is the answer and the comparison is already over — it is priced, available, and routable, and its published 720p rate of $0.084 per second makes the entry tier affordable to iterate against. Its multi-shot mode is a genuine workflow advantage over single-clip generation, and its known weaknesses in multi-object physics matter less for the work most teams are actually doing, which is short-form video where the viewer will not run a physics check.

If you need a controller that works on hardware, Odyssey-3 is the one pointed at your problem and there is still nothing to buy. The concrete reason to care is the breadth: one set of weights reaching arms, humanoids, a car, a drone, and three games is a claim about generality, and generality is what makes a model a foundation model rather than a policy. The concrete reason to wait is that no one outside Odyssey has run it. Watch for a price and an independent evaluation, in that order.