Daily AI Brief, Oct 28: Genie 4 Arrives Gated as Gemini Robotics ER 2 Opens in Preview

Google put two physical-AI models into preview inside twenty-four hours this week, and it gated them in opposite directions. Genie 4, the Google DeepMind world model that generates a steerable scene with synchronized sound, was announced on October 27, 2026 for a limited group of trusted testers, with a waitlist and no published price. Gemini Robotics Embodied Reasoning 2 — ER 2, the high-level robot planner Google Cloud moved to public preview the same day — comes with a model ID, a documentation page and a sign-up any Cloud customer can complete. Both are previews. Only one of them has a door you can walk through today.

That asymmetry is the story of the week, more than either model on its own. Google is assembling the two halves of a physical-AI stack: a world model that can manufacture experience, and an embodied reasoning model that can plan against real robots. The halves were announced a day apart and are being distributed by two different parts of the company under two different access regimes. What follows is a brief, not a review — every figure below is labeled with who produced it.

Item 1 — Genie 4, and the three claims doing the work

The DeepMind post is short on numbers and long on capability claims. Three of them matter:

• Native audio. Genie 4 generates ambient sound, music and interactive dialogue in the same pass as the frames, rather than dubbing a silent simulation afterward. It is the first DeepMind world model to do this.

• Live steering. A typed instruction can move the camera, change the weather or introduce a character while the world is running, and a Gemini model adapts the instruction to what is currently on screen.

• A longer window. The figure in circulation is roughly two minutes of scene consistency, double Genie 3. That number came from launch briefings, not from the blog: Google's own text says the model sustains "short sessions" and does not put a duration in writing. Treat two minutes as a reported figure with a vendor behind it, not a specification.

The structural addition is easier to miss than the audio. Gemini Omni anchors the simulation in real-world appearance and behavior, and Gemini 3.8 Flash sits in the loop, watching generated frames, breaking a goal into steps and re-planning when the scene diverges. In the demo Google chose to publicize, Gemini 4 Argon decomposed a task that Genie 4 then simulated in real time, with an Apptronik Apollo humanoid playing checkers. That is a different product from a video generator: a generator answers a prompt with a clip, while a generator plus a planner answers an objective with a sequence of checked states.

What Google did not publish is as consequential as what it did. There is no resolution, no frame rate, no parameter count, no latency measurement, no benchmark score and no price. Every "Genie 4 is more realistic" comparison published this week is a comparison of demo reels and feature lists, because nobody outside the trusted-tester pool has measured output. Google does concede the failure modes: detail and object permanence drift in longer sessions and after repeated edits, spatial accuracy is imperfect, in-world text is often illegible, and audio accuracy remains unsolved. All generated video and audio carries a SynthID watermark.

Item 2 — Gemini Robotics ER 2, which you can actually sign up for

ER 2 is not new this week as a model — Google introduced it on July 30, 2026 in a limited private preview — but the October 27 announcement is the one that changes who can use it. It is now in public preview on the Gemini platform through Model Garden, and tryable in Agent Studio, with the model ID gemini-robotics-er-2-preview.

Google positions it as a robot's high-level brain rather than a controller. It takes spatial and temporal video understanding, long-horizon planning and safety guardrails and wraps them in an orchestration layer: developers declare low-level control interfaces — vision-language-action models, navigation APIs, custom functions — as tools, and ER 2 plans over them. It also classifies task progress, which is what lets a robot notice a step failed and retry it without restarting the whole workflow.

The vendor-reported figures Google leads with:

• Event-frame pinpointing — 91.3% accuracy at identifying the exact frame where a critical event occurs, with a 0.96-second mean absolute frame distance.

• Execution speed — 4x, described as "a fraction of the compute cost of many larger model categories." No baseline is named in the sentence, so the multiple is directional rather than comparable.

• Instrument reading — 10+ categories including circular dials, sight glasses, digital displays, linear scales and liquid thermometers, read directly from visual input.

• Safety — 97.9% on Safety Instruction Following and 93.0% on Human Proximity at a one-meter threshold, per Google's own safety reporting. Physical protections such as e-stops, barriers and force limits remain the operator's responsibility; the scores describe reasoning, not hardware.

There is no rate card in the announcement itself, only a link to the pricing page. For a preview model that is normal and it is also the number a procurement team needs first.

The pattern behind the two items

Read together, the announcements describe a stack that is being built in public and sold in private. A world model is only useful for robot training if it can generate enough varied, reset-able experience to matter; a robot planner is only useful if it can be evaluated against something other than a fixed video corpus. Google is clearly pointing the two at each other. It is equally clearly not ready to hand either one to an unvetted customer: Genie 4 is tester-only, ER 2 is preview-only, and neither has a general-availability date.

That leaves a practical question for anyone building on the stack: which parts are callable now?

• Not callable. Genie 4, until the waitlist opens or an early-access partner agreement lands. Gemini Robotics ER 2, unless you are prepared to build against a public preview on Google Cloud.

• Callable today. Gemini Robotics ER 1.6 preview is in OrcaRouter's catalogue, and so is Gemini 3.8 Flash — the same model Google names as the in-loop agent that checks Genie 4's frames and re-plans. Neither is the world model. Both are the reasoning half of the pipeline, and they can be exercised now rather than after a waitlist.

OrcaRouter routes those through one API across 200+ models with provider list prices passed through at 0% markup, so a Google price change reaches our catalogue the same day it is published. Automatic failover matters more than usual here: a preview-tier model can be rate-limited or withdrawn, and a second provider behind the same call is the difference between an experiment and an outage. The routing DSL is the other option when a pipeline needs the planner and a separate model for verification in one composed call. We do not host Genie 4, and nothing in this brief should be read as a route to it.

What would change this brief

Four events would move Genie 4 out of the "demo" column and into something a team can budget against: a published spec sheet with resolution and frame rate; a price for the Gemini API tier; a first independent measurement from outside the trusted-tester pool; and Project Genie shipping Genie 4 to AI Ultra subscribers, which is the only consumer surface Google has named. The equivalent event for ER 2 is a rate card and a general-availability date.

Until then, the honest summary is that Google has announced the most capable world model it has built and the most usable robot planner it has shipped, in the same week, and put the more impressive of the two behind a waitlist. The one you can call today is the smaller announcement. That is usually how preview weeks go.