Genie 4 vs Odyssey-3: A Measured World Against an Unmeasured One

Google DeepMind announced Genie 4 on October 27, 2026, as a research preview for trusted testers. Odyssey opened a free browser preview of Odyssey-3 on October 8, 2026, three weeks after announcing it. Put the two side by side and one asymmetry decides most of the argument: Odyssey-3 has a row on a third-party leaderboard that anyone can go and read, and Genie 4 has no published number of any kind — no resolution, no frame rate, no latency, no benchmark, no price. That does not make Odyssey-3 the better model. It makes Odyssey-3 the only one of the two you can currently form an opinion about without taking a vendor's word for it.

Both are interactive world models. Both generate video of a scene rather than 3D geometry you could export. Both are aimed, explicitly, at the same commercial job: giving robots and agents a place to practise. And both are walls of caveats dressed as products. The interesting comparison is not which one renders a more convincing meadow. It is what each company chose to expose, and what that choice tells you about how close either is to being usable.

What each one actually is

Genie 4 is the fourth generation of DeepMind's world simulator, following Genie 3 in August 2025. It generates an explorable scene from a prompt, lets you steer it with typed instructions while it runs, and now produces audio in the same pass as video. Google calls it the first real-time world model to generate audio natively. A Gemini model sits in the loop as a planner, checking the frames against the goal and re-planning when they diverge.

Odyssey-3 comes from Odyssey, a Palo Alto lab founded by self-driving veterans that raised a $310M Series B in June 2026 at a $1.45B valuation. It is an autoregressive diffusion transformer trained to predict how a scene evolves in response to actions — an action-conditioned dynamical system rather than a video generator with a text box. You drive it first-person, third-person or with a free camera, and it responds to movement the way an environment does.

The dimensions that matter, side by side:

• Coherence window — Genie 4 holds a scene for about two minutes, double Genie 3's; Odyssey-3 publishes no session length at all.

• Resolution — Genie 4 has published none; Odyssey-3 ships at 832×480 on the base tier and 1280×720 on Odyssey-3 Pro.

• Frame rate — unpublished on both sides, though a third-party Physics-IQ column lists 16 fps for Odyssey's entries.

• Audio — native ambient sound, music and interactive dialogue in Genie 4; absent from Odyssey-3.

• Steering — natural-language edits to weather, camera and objects mid-scene in Genie 4; movement and camera control in Odyssey-3, without a language director.

• Benchmark — none published for Genie 4; one third-party submission and one self-run board for Odyssey-3.

• Access — trusted testers and a waitlist for Genie 4; a free browser demo today for Odyssey-3, with API access by request.

The number that separates them

Physics-IQ Verified is the closest thing this category has to an independent scoreboard. It is run with Anates Labs and DeepMind, and it scores whether a generated scene obeys physical law when something is dropped, poured, swung or struck. Odyssey submitted Odyssey-3 and Odyssey-3 Pro to it. Google submitted Genie 4 to nothing.

On the video-to-video track, read on October 13, 2026, Odyssey-3 Pro scored 66.10 and ranked first. Base Odyssey-3 scored 64.43. Those are the figures Odyssey quotes, and they come with a sampling condition that changes how you should read them: they are best-of-eight runs, where the model gets eight attempts and the best one is scored. On a single generation the same models score 63.37 and 61.56. Give Odyssey-3 the benchmark's own prompts instead of prompts it wrote for itself and the score falls to 51.76, eighth place, roughly ten points below its headline.

That is the most useful thing on the board, because it is the closest approximation of what a customer would actually experience. A team deploying this model does not get eight rolls and a scorer. It types a prompt and gets one world. The honest read is that Odyssey-3's physics is genuinely strong relative to video generators — it beats FLUX 3 [large] on the video-to-video track — and that a meaningful part of its lead is selection effect. Odyssey's other claim, first place in three of four WorldMark categories, is weaker still: that evaluation was run by Odyssey, Odyssey-3 does not appear in WorldMark's published results, and Odyssey cites thirteen metrics where the board describes nine.

Genie 4's row on that table is blank, and the blank is not neutral. Google shipped a launch with an unusual hole in it. Genie 3's reveal material in 2025 was shown at 720p and 24 frames per second; Genie 4's blog post discusses realism without giving either number. When a lab publishes a quality claim and withholds the two figures that would let you check it, the withholding is the message. Something about the current output does not survive being written down next to its predecessor.

Two different bets about where the intelligence belongs

The architectural split is more consequential than the spec sheet, because it determines what you can build on top of each model.

Genie 4 ships a pilot. Gemini 3.8 Flash watches the generated frames, decomposes an objective into steps, judges whether the scene did what was asked, and revises the plan. Gemini Omni anchors the simulation in how things actually look and behave. In one experiment, Gemini 4 Argon broke a robotics task into captions that Genie 4 then simulated in real time, including an Apptronik Apollo humanoid playing checkers. The claim underneath all of that is that the model can rehearse a task on its own, without a human driving every frame.

Odyssey-3 ships a cockpit. There is no planner in the loop. The model is conditioned on your actions, and the quality of the session is a function of how well you drive it. That is a smaller claim, and it is a more honest one about what the technology currently does — a world model that needs a competent operator is still useful for evaluation, data generation and exploration, none of which require autonomy.

The practical consequence: if your use case is "generate diverse failure cases for a policy to learn from," Odyssey-3's shape is closer to what you need and you can try it today in a browser. If your use case is "an agent that sets its own curriculum," Genie 4 is the only one of the two making that claim, and nobody outside the tester pool has verified it.

Robotics is the job both are auditioning for

Neither company is coy about the target. Google frames Genie 4 around training environments for robots and agents. Odyssey demonstrated a robotic arm recovering from manipulation errors, humanoid control through Flexion policies, a car driving real roads in India after roughly twenty hours of training data, and indoor drone navigation — six domains from one base model, all vendor-reported and none independently reproduced.

The claim worth taking seriously is cross-embodiment transfer, because it is the one that would make a world model economically different from a simulator. If dynamics learned from one body transfer to another, you stop hand-building physics for every robot. Odyssey says it demonstrated early transfer between games, playing one title with skills learned in another. That is a long way from a factory floor, but it is the right direction, and it is the kind of result a competitor would want to replicate rather than dispute.

Genie 4's robotics evidence is a demo of a humanoid playing a board game in simulation. It is charming and it proves the pipeline connects. It does not prove the simulation is faithful enough to train on, and with no resolution, frame rate or physics score published, there is no way to check.

Audio is real, and it is not a tiebreaker

Genie 4 generates ambient sound, music and spoken dialogue that responds to what you say to a character in the scene. It is a genuine first for DeepMind's world models and it changes which jobs the model can plausibly do — a pre-visualisation shot without sound is half a shot, and a sound designer who has to rebuild the room tone from scratch is working from the model's timing rather than their own.

It also does not decide this comparison. Odyssey-3 is not trying to be a film tool; it is trying to be a physical environment. Silence costs it nothing in a robotics curriculum and costs it a great deal in a storyboard review. Google's own caveats undercut the feature anyway: the company lists real-time audio accuracy as unsolved, alongside legible in-world text and spatial accuracy. A world whose signage you cannot read is not a stand-in for a street, and dialogue that drifts from its source is not something you can cut against.

What access actually costs

You can open Odyssey-3 in a browser this afternoon and drive it for free. The public demo runs a distilled few-step variant the company calls Odyssey-3 Flash — a third name Odyssey does not otherwise document — while the full models sit behind a priority-access request. There is no self-serve API, no price list, no weights, no licence, and the only terms page on the site predates Odyssey-3 and still describes the Odyssey-2 prototype.

Genie 4 is gated harder. Trusted testers have it now; Google AI Ultra subscribers are promised access through Project Genie "in the coming weeks" at roughly $100 a month, with early-access partners reaching it through the Gemini API and AI Studio. There is no general-availability date and no developer price. One outside estimate circulating in the Chinese-language press put a two-minute session at about $7.84 in compute — single-sourced, unconfirmed by Google, and not a price anyone has been quoted.

That asymmetry is the whole procurement story. Odyssey-3 costs you an afternoon to evaluate and tells you nothing about what it will cost to run in production. Genie 4 costs you a waitlist and tells you nothing about anything. If you need an answer this quarter, only one of these questions is even askable.

How to choose this quarter

Start with what you can measure. If your need is an interactive environment for exploration, data generation or a demo, try Odyssey-3 in the browser, then ask Odyssey what a session costs and what the concurrency limits are — the answers will tell you more about the product's maturity than any benchmark. Treat its 66.10 as a best-of-eight number and its 51.76 as the realistic one, and plan around the second.

If your need is an agent that plans inside a simulation, Genie 4 is the only option making that claim, and there is nothing to test. The rational move is to wait for the waitlist and, in the meantime, build against the half of the architecture that is already a commodity: the planner. Gemini 3.8 Flash is the model Google put in Genie 4's loop, and it is a routable LLM today rather than a gated world model. It is on our catalogue as google/gemini-3.8-flash, callable through one API with provider list pricing passed through at 0% markup. OrcaRouter does not host Genie 4 or Odyssey-3 — neither is an addressable endpoint anywhere — so the honest split is a planner you can buy and a simulator you cannot yet. In that shape, failover matters: if the only swappable part of your stack is the guiding agent, you want it to move between providers without a new contract, and the routing DSL lets you pin a cheap model for verification passes and a stronger one for planning.

There is also a cheaper experiment nobody is running. Point the same planner at Odyssey-3's browser session and at a conventional physics engine, and see which one produces better training data per dollar. That comparison costs a week and would settle more than either launch post did.

What would settle it

Three things, in order of usefulness. A resolution and frame rate for Genie 4, from Google or from any tester willing to publish a capture at native size. A matched-harness run of both models on Physics-IQ with benchmark-provided prompts, single generation, which would put the same number on both rows for the first time. And commercial terms — a per-second rate, a licence, a concurrency figure — because a world model you cannot price is a demo, whichever lab built it.

Until one of those lands, the accurate summary is uncomfortable for both companies. Odyssey-3 is measurable and narrow, with one strong third-party result that shrinks under realistic sampling. Genie 4 is broader and unmeasurable, with a planner inside it and no evidence yet that the world it plans against is faithful. The gap between those two positions is not a gap in capability. It is a gap in disclosure, and it is the reader's job to notice which one is being asked to take it on faith.