Genie 4 vs Genie 3: Two Minutes, One New Sense, and the Same Locked Door

Genie 4 was announced on October 27, 2026, 448 days after Genie 3. The interval is easy to state and hard to evaluate, because the two models are not sold the same way and, strictly speaking, neither is sold at all. Genie 3 has been a gated research preview since August 5, 2025, reachable by Google AI Ultra subscribers through Project Genie since January 2026. Genie 4 is a gated research preview again, this time for a limited set of trusted testers with a public waitlist. What changed in between is real: sound, live steering, a longer coherence window and a second Gemini model watching the output. What did not change is the part that decides whether you can use it.

Most Genie 4 coverage reads as a list of new features, which is accurate and also incomplete. The more useful comparison is between two generations of the same idea at two different stages of maturity: Genie 3 is a known quantity with published resolution and frame rate and a product path, and Genie 4 is a demo with a better description and no specification sheet. That does not make Genie 4 vapor — the blog, the demos and the robotics experiment are real — but it does mean the honest comparison has two columns of different kinds of evidence.

The four additions, ordered by how much they change the job

Not every new capability matters equally. Ranked by what they do to a pipeline:

• Native audio. This is the change that alters the product category. A silent simulation can drive a camera decision but cannot stand in for a shot, because the sound designer still has to build rain and footsteps against timing the model did not produce. Genie 4 generates ambient sound, music and interactive dialogue in the same pass as the frames, which means the audio and the picture share a clock. Google's caveat is important: audio accuracy is unsolved, and generated dialogue is the least predictable part of the release.

• Live steering. Genie 3 could accept promptable world events — change the weather, add an object — but the interaction was coarse. Genie 4 accepts a typed instruction mid-simulation and a Gemini model adapts it to the current frame, so the camera can move and the scene can change without a reset. For pre-visualization this is the difference between generating a shot and directing one.

• The longer window. The number everyone quotes is roughly two minutes of scene consistency, double Genie 3. It is a reported figure from launch briefings; the blog says only that sessions are "short." The practical effect is that a take can last long enough to block a short sequence, not that a level can be explored.

• The agent in the loop. Gemini Omni grounds appearance and behavior, and Gemini 3.8 Flash watches frames, decomposes goals, verifies progress and re-plans. That is an architectural change rather than a rendering one, and it is why the robotics demo works: Gemini 4 Argon planned, Genie 4 simulated, and an Apptronik Apollo humanoid played checkers in the generated world.

What Genie 4 did not publish

This is where the comparison stops being symmetric. Genie 3 has a public specification: 720p resolution, 24 frames per second, real-time interaction, visual memory reaching back about one minute, and sessions lasting a few minutes. Genie 4 has none of those numbers in writing. No resolution, no frame rate, no latency, no model size, no benchmark, no price.

That absence has a specific consequence for anyone tempted to write "Genie 4 is higher fidelity than Genie 3." It might be true. The demo reels look better, and Google claims improved physics, more natural human motion and more photorealistic scenes. But without a published resolution or frame rate, there is no way to tell whether the improvement is the model or the capture. A demo rendered at a higher resolution than the product ships at would look exactly like a better model. Every fidelity claim in this comparison is vendor-reported and unreproduced, and the specification gap is the reason to say so out loud rather than in a footnote.

The admitted limitations are also more specific this time. Google concedes that detail and object permanence drift in longer sessions and after repeated complex edits, that spatial accuracy is imperfect, that in-world text renders badly, and that audio accuracy is a research problem. Those are the same categories Genie 3 listed — constrained action space, weak multi-agent interaction, poor real-world geography, legibility — with audio added. The list did not shrink. It grew a category.

The door did not move

Genie 3's access story is a product story. Project Genie, the Google Labs surface released on January 29, 2026, gives AI Ultra subscribers a way into Genie 3 worlds, and it was reported to cap a single world exploration at 60 seconds — a compute decision, since autoregressive generation gets more expensive the longer a session runs. That cap is worth holding in mind when reading the two-minute figure for Genie 4: the model's coherence window and the product's session limit are two different numbers, and the product number has historically been the tighter one.

Genie 4's access story is a preview story. Trusted testers now, AI Ultra subscribers through Project Genie "coming soon," early-access partners through the Gemini API and AI Studio "coming soon," and a waitlist in the meantime. There is no date attached to either "soon." For a developer, that means the upgrade is not a purchasing decision yet. For an AI Ultra subscriber, it means the model you can reach today is still Genie 3.

The comparison a reader can act on is therefore narrow. If you already have Project Genie access, Genie 4 changes nothing about your week until it ships there. If you are choosing what to prototype against, the reasoning layer is the part that exists: Gemini 3.8 Flash, the model Google names as Genie 4's in-loop agent, is callable now through OrcaRouter on the same key as Gemini Robotics ER 1.6 preview, with provider list prices passed through at 0% markup and automatic failover across providers. That lets a team build the planner half of a simulation pipeline before the world model is available to them. It does not substitute for Genie 4, and we do not host either Genie model.

What would settle it

Three publications would turn this from a feature comparison into a model comparison. The first is a spec sheet: resolution, frame rate, latency and session limit, in writing, from Google. The second is an independent measurement from outside the trusted-tester pool — anything that reports frame rate under load or coherence time under repeated edits, rather than repeating the briefing figure. The third is Project Genie's session cap for Genie 4, which will tell you what the two-minute claim is worth in the product people can actually buy.

Until then, Genie 4 vs Genie 3 is a comparison between a shipped preview with known limits and an announced successor with better demos and no specs. The successor is probably the better model. The predecessor is the only one anyone outside Google has used.

Who should care now, and who should wait

Researchers already inside a tester program should care now: audio plus steering plus an agent loop is a materially different experimentation surface, and the robotics demo suggests where Google wants the model pointed. Studios doing pre-visualization should wait for two things — Genie 4 landing in Project Genie, and a licensing answer for generated audio, which the launch material does not provide. Game teams should wait for the session limit, because a two-minute coherence window is a shot, not a level, and the failure mode Google itself lists — object permanence drifting after repeated edits — is exactly the one that breaks a walkthrough.

None of that makes Genie 4 a disappointment. It makes it an announcement, which is what it is. The gap between the two generations is real and, on the evidence available, larger than the gap between Genie 2 and Genie 3. The gap between owning either of them is zero.