Title card headed Genie 4 with the line 720p at up to 24 fps, about a minute per session, and three cards for real-time 720p, edit mid-session and native audio, footnoted to Google's model card of 18 November 2026.
Guides & Insights

Genie 4 Is Live at 720p and 24 Frames a Second, and That Number Is the Story

Author

Elias Hawthorne

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

The most informative line in G​oogle's 18 November Genie 4 release is not in the blog post. It is in the model card attached to it: real-time 720p video at up to 24 frames per second with synchronized audio, generated from a text prompt, an image, or live keyboard and mouse input. For three weeks Genie 4 had been a set of demos with no specification at all — no resolution, no frame rate, no session length. The Project Genie rollout to G​oogle AI Ultra subscribers did not just open a door; it published the first hard numbers, and those numbers describe a much smaller thing than the launch reel suggests, in a way that is actually useful if you are deciding what to do with it.

The other half of the release matters too, and it is the half G​oogle led with: mid-session text editing, native audio that adapts as a scene unfolds, a conversational agent that builds worlds, and a community gallery with remixing. Genie 4 is the model; Project Genie is now the product around it. Separating the two is what this piece does, because the model card and the gallery answer different questions — one tells you what the system is, the other tells you what G​oogle thinks you will do with it.

What actually changed on 18 November

G​oogle announced Genie 4 on 27 October 2026 as a research preview for trusted testers with a public waitlist. Nothing about that model changed on 18 November. What changed is who can reach it and what sits around it.

• Access — Genie 4 rolls out to eligible G​oogle AI Ultra subscribers inside Project Genie. G​oogle AI Pro members can join a waitlist. There is still no API model ID, no per-generation price, and no general-availability date for developers.

• Real-time editing — a typed prompt mid-session changes weather, adds a character, or redirects the action without restarting the world. G​oogle's demo of this is a hillside cafe turned into a glass pavilion in a desert.

• Native audio — ambience, music and character dialogue generated alongside the frames rather than layered on afterwards. This is the first Genie that makes sound at all, and G​oogle's own docs list real-time audio accuracy as an open research problem.

• The agent — a conversational builder that orchestrates G​oogle's multimodal models to establish a world's look and a custom score, then keeps working during the session: a G​emini model watches the screen, offers suggestion chips for where to go next, and can trigger in-world events.

• The gallery — curated worlds plus community uploads, all remixable. Two named collaborations ship with it: "Villa Giallo," a room-by-room gothic-manor mystery built with creative technologist Will (BLVCKL!GHT), and "The Life of Ants," an ant's-eye colony walkthrough with filmmaker Anand Gandhi.

• Genie 3 got the upgrades — and this is the part that will annoy anyone who skimmed the headline. The older model now runs up to three minutes per generation, up from sixty seconds, supports world chaining so several worlds link into one continuous journey, accepts 360-degree panorama images as seeds, and takes real-time in-world event prompts. Genie 3 remains in the model picker for exactly those cases.

The spec sheet, and the one constraint it names out loud

Before 18 November, any claim about Genie 4's fidelity was unfalsifiable. The launch post said "most realistic and interactive world model yet" and showed clips; it published no resolution, no frame rate, no latency figure. The November model card closes part of that gap:

• Output — 720p, up to 24 fps, with synchronized audio, generated in real time.

• Inputs — text, images, and live user actions: keyboard and mouse navigation, not just a prompt.

• Session length — "around a minute of continuous interaction" on the model page's limitations list, with the model card saying only that it holds coherent environments over "short sessions."

• Base — built on G​emini Omni Flash, with G​emini models supplying reasoning and real-time steering.

• Constraints G​oogle admits — visual detail and object permanence drift over longer runs, after several complex prompt edits, or while moving through a scene; complex physics, audio accuracy, legible text, and spatial accuracy are all listed as unsolved.

• Provenance — SynthID watermarking on every frame of video and every second of audio.

Generated six-row scoreboard for Genie 4 listing 720p at up to 24 fps, text image and live keyboard and mouse inputs, roughly one-minute sessions, Gemini Omni Flash as the base, known limits of drift physics and legible text, and SynthID on video and audio, with a footer noting no independent benchmark exists.

Read those two lines together — 24 fps at 720p, and about a minute — and the honest description of Genie 4 is a real-time renderer running at roughly the quality of a 2010s game console, for as long as it takes to walk down a corridor and back. That is not a criticism. Real-time generation at a stable frame rate is a far harder problem than producing a beautiful eight-second clip, and 720p at 24 fps is the resolution and cadence of broadcast television, not of a toy. But it does mean the launch reels are showing you the ceiling of a very short session, and the drift G​oogle documents is what happens when you try to push past it.

Why a minute is a product decision, not a limitation to apologise for

A minute is enough for a shot, a room, a walkthrough, a single interaction — the unit of work in pre-visualisation, spatial design review, and interactive teaching. It is not enough for a level, a film sequence, or a training episode that needs twenty minutes of consistent state. G​oogle's answer to that gap is not a longer Genie 4; it is Genie 3 at three minutes with world chaining, sitting next to it in the same picker. That is an unusual product decision — the newer model is not strictly the more capable one for every job — and it tells you the two are being positioned by session length rather than by generation.

The gallery collaborations read the same way. "Villa Giallo" is a mystery you solve room by room, which is a structure built out of one-minute units. "The Life of Ants" is a perspective piece — you navigate a colony from an ant's eye view — which needs atmosphere and scale more than it needs continuity. Both are demos chosen to show what short interactive worlds are good at, and neither pretends to be a film.

For a studio, the practical question is where a one-minute interactive take fits in a pipeline that currently runs on eight-second generated clips. The answer is upstream and sideways: blocking, camera language, set dressing decisions, and playtests of spatial ideas — the things you would otherwise build in a grey-box engine and throw away. Genie 4 removes the grey box. It does not remove the engine.

What you can call today, and what you cannot

Genie 4 is not an endpoint. It is a browser product behind a $99.99-a-month G​oogle AI Ultra subscription, with no published API, no model ID, and no per-generation rate. That is true of every third-party platform as well, ours included: there is nothing to route to.

What is callable is the reasoning half of the architecture G​oogle describes. The launch material names Gemini 3.8 Flash as the in-loop agent that plans steps and checks frames against goals, and that model is a live endpoint on our catalogue as google/gemini-3.8-flash at $0.75 per million input tokens and $3.75 per million output — provider list price passed through with no markup on top, so a G​oogle price change reaches your bill the same day it reaches G​oogle's rate card. If you are building the loop around a world model rather than the world model itself — planning, verifying, re-planning, writing the prompts that steer a session — that is the piece you can integrate this week, with automatic failover across providers behind one key.

Screenshot of the OrcaRouter model page for google/gemini-3.8-flash, showing the model identifier, its 65K maximum output and per-million-token pricing fields.

What would move this from a product to a platform

Three things, in ascending order of how much they would change the answer.

An API. A model ID and a per-second or per-session price would let Genie 4 compete for pipeline slots against Veo, K​ling, and Seedance on the same spreadsheet. Without one, it competes for a line in a creative-tools budget.

An independent measurement. There is no benchmark, no Elo, no human-preference study, and no third-party frame-rate analysis of Genie 4 anywhere. Every fidelity claim in circulation traces back to G​oogle's own reels. The 720p and 24 fps figures are now checkable against the model card, which is progress; the claim that it is more realistic than Genie 3 still is not.

Session length past a minute at Genie 4's quality. G​oogle is currently solving continuity by selling you the older model. When Genie 4 closes that gap, the picker collapses back to one option and the interesting comparison stops being Genie 4 vs Genie 3 and starts being Genie 4 vs everything else.

Screenshot of Google DeepMind's Genie 4 announcement dated 27 October 2026, headed Genie 4: A new generation of world simulation, crediting Raia Hadsell, Diego Rivas and Hang Qi, with a Join the Genie 4 waitlist button.

Until then, the fair summary of 18 November is that G​oogle shipped a creative ecosystem around a model it had already announced, published the model's first real specification, and left the one number developers care about — the price of a second of output — unpublished. For a $99.99 subscription you get the ecosystem, the gallery, the agent, and a minute at a time. Whether that is a bargain depends entirely on whether your work fits in a minute, and no one outside G​oogle can tell you that yet.