As of October 29, 2026, Google DeepMind's model portfolio covers five distinct units of work: general reasoning, video generation, music generation, interactive world simulation, and robot control. Gemini 4 Argon and Genie 4 entered the current map in October, and Google now documents several cross-model integrations. That still does not make every product a lossless production pipeline.
The durable way to choose among them is to define inputs, outputs, latency, safety, and human-rework requirements first. Model names will keep changing. Workflow constraints change more slowly.
Update note: This revision preserves the original URL and reader task while correcting unsupported architecture, duration, context-window, and ecosystem claims. Version and pricing statements are anchored to official pages checked on October 29, 2026.
The short version: five models, five units of work
| Unit of work | Current representative | What official sources establish | What you still need to test |
|---|---|---|---|
| Complex reasoning and multimodal understanding | Gemini 4 Argon | Current flagship for complex software, enterprise-knowledge, and cybersecurity workflows | Account access, real-task quality, latency, long-context degradation, and cost |
| Video generation and editing | Veo 3.1 | Native audio, reference controls, scene extension, object editing, 1080p and 4K output | Endpoint lifecycle, continuity, speech quality, and cost |
| Music generation | Lyria 3.5 | Tracks up to three minutes, lyrics and vocals, image-to-music, SynthID | Licensing, editability, style consistency, and mixing work |
| Interactive world simulation | Genie 4 | Multimodal worlds, ambient sound, music, and real-time dialogue; currently a trusted-tester research preview | Access, action latency, controllability, and task fitness |
| Physical-world reasoning and control | Gemini Robotics 2 / ER 2 | ER handles high-level understanding and planning; VLA handles action | Hardware adaptation, safety constraints, recovery, and production access |
There is deliberately no overall winner. Reasoning, video, music, simulated worlds, and physical action have different output spaces. Compressing them into one leaderboard removes the engineering question that matters.
Gemini: Argon is current; 3.1 Pro remains a live preview
Google released Gemini 4 Argon in October 2026. Its model card positions it as the frontier reasoning model in the Gemini 4 family for software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense. It has replaced Gemini 3.1 Pro as the current flagship, so a new evaluation should begin by checking actual Argon access through the relevant account and API.
Gemini 3.1 Pro has not disappeared. Google's deprecation table still lists gemini-3.1-pro-preview with no announced shutdown date. Its model card confirms text, image, audio, and video input, up to a 1M-token input context, and 64K text output. The architecture section says that it is based on Gemini 3 Pro. It does not disclose an expert count, routing policy, or MoE configuration that would justify a more specific public claim.
In Google's February 2026 evaluation, Gemini 3.1 Pro Thinking High scored 77.1% on ARC-AGI-2 and 94.3% on GPQA Diamond. Those numbers answer a narrow question: which evaluations Google uses to support the model's positioning. They do not independently prove that the model wins on your repository, report, or agent workflow. Vendors, reasoning settings, and evaluation harnesses all affect results.
At the time of this review, Google's pricing page lists standard gemini-3.1-pro-preview pricing below 200,000 input tokens at $2 per million input tokens and $12 per million output and thinking tokens; above that threshold it is $4 and $18. That is a useful cost baseline for a live legacy integration, not Argon pricing. Pricing, preview status, and model IDs are highly time-sensitive and should be checked again before deployment.
The useful acceptance test is a frozen set of real tasks: long-document questions, cross-file edits, tool calls, and structured-output constraints. Measure task success, human repair time, end-to-end latency, and cost per accepted result instead of copying a public leaderboard.
Veo 3.1: validate video by entry point and shot
Veo 3.1's official product page lists native audio, reference-image controls, first-and-last-frame control, scene extension, outpainting, object insertion and removal, character and motion controls, plus 1080p and 4K output. Product surfaces and API lifecycles must be separated: the three Veo 3.1 Preview endpoints in the Gemini API shut down on October 22, 2026. Google recommends Gemini Omni Flash or the Veo 3.1 GA models on Gemini Enterprise Agent Platform. Consumer surfaces such as Flow and Gemini require separate account-level verification.
Multiple entry points do not guarantee identical limits. Google's public comparison material primarily evaluates six-to-eight-second clips, while the product page separately offers scene extension. The defensible statement is that Veo can generate and extend short shots, not that an unsupported 90-second number is the universal single-generation limit.
Google also lists natural and consistent short spoken audio as an area under active development. A production test should therefore cover character continuity, shot transitions, audio-video synchronization, speech intelligibility, physical plausibility after object edits, and SynthID handling. A 4K checkbox is not enough.
Lyria 3.5: from generation to directed creation
DeepMind's current music model at this review date is Lyria 3.5. Its model card says it uses latent diffusion over temporal audio latents, accepts text, and outputs music audio and lyrics. The product page supports tracks up to three minutes, lyrics and vocals, duration prompting, image-to-music, and SynthID watermarking.
These features cover the path from a creative description to high-fidelity audio. Professional production still depends on structural editability, export formats, licensing, repeatability, and human repair in a DAW. The current official pages do not support the previous article's MIDI-output claim, so this revision removes it. lyria-3-pro-preview remains an officially listed previous-generation preview model; the current Lyria 3.5 model ID is lyria-3.5. They should not be conflated.
For the earlier quality trajectory, see How DeepMind Pushed Past AI Music's Quality Ceiling.
Genie 4: the current world model, with Genie 3 as a baseline
DeepMind announced Genie 4 on October 27, 2026. The official post describes a new multimodal world model that can generate environments, music, ambient sound, and interactive dialogue together. Current access is a research preview for trusted testers, with rollout beginning through Project Genie for Google AI Ultra subscribers and early partners of the Gemini API and AI Studio.
Genie 3 remains a useful historical baseline. Its 2025 announcement established text-generated navigable worlds at 720p and 24 frames per second with consistency for a few minutes. It should no longer be listed as the current October 2026 version.
The distinction from Veo is the interaction loop. A video model delivers media. A world model must generate the next state after every user action. Acceptance therefore shifts to action latency, state persistence, controllability, run-to-run variation, and coverage of the target task.
Both generations are presented as research capabilities. Treating Genie 3's 720p and 24 FPS, or Genie 4's real-time demonstrations, as a production simulation SLA would exceed the available evidence.
Gemini Robotics 2: separate planning from action
The Gemini Robotics 2 family, announced in July 2026, includes Gemini Robotics 2, Gemini Robotics ER 2, and On-Device 2. Google describes ER 2 as a high-level brain that understands physical space, plans multi-step tasks, uses tools, and detects completion before handing motor execution to a VLA model. The ER 2 model card says it is based on Gemini 3.5 Flash, accepts text, image, video, and audio, has up to 128K input context, and produces up to 64K text output.
That separation matters more than one benchmark number. High-level planning plus lower-level control creates additional failure surfaces: connectivity, action authority, state drift, and emergency stops. Google's model card also urges caution in production, commercial, or public environments and excludes safety-critical work such as healthcare and transportation.
The earlier article's ER 1.6 instrument-reading result belongs to a specific April 2026 evaluation. It documents progress, but it cannot replace acceptance testing of ER 2 on the target hardware. For the earlier generation, see Gemini Robotics and AI in the Physical World.
How much of an ecosystem is publicly established?
Public sources support three forms of commonality: the models come from Google DeepMind; several use Google TPUs, JAX, or ML Pathways; and multiple products are distributed through Gemini, AI Studio, Flow, or Google Cloud.
The Genie 4 announcement also documents specific integration: Gemini Omni anchors real-world appearance and behavior, Gemini 3.8 Flash acts as an in-loop guiding agent, and Google is experimenting with Argon, Genie 4, and Robotics together. Those are concrete model-level integration claims.
The boundary still matters. Those named combinations do not establish that all five families share one learned representation or can pass arbitrary outputs among one another without loss. Shared distribution, shared infrastructure, and documented model integration are different claims.
Ecosystem value must therefore be proven in a workflow. If Gemini produces a Veo storyboard, measure prompt-translation loss, character consistency, and human rework. If a Genie environment is used for robot training, measure the simulation-to-real gap. Brand ownership is not interface validation.
A selection framework that ages better than a model list
Freeze the task before comparing model names.
- Define the unit of work: reasoning, media, an interactive environment, or physical action.
- Build acceptance samples from real inputs, including normal, boundary, and failure cases.
- Count the full cost: API usage, latency, human rework, format conversion, and safety review.
- Separate evidence levels: product pages describe capabilities, model cards describe boundaries, vendor evaluations report vendor results, and internal trials establish workflow fit.
- Set revalidation triggers: rerun tests when model IDs, preview status, pricing, access paths, or critical limits change.
The goal is not to make an article permanently correct. It is to make clear which decisions must be retested when the environment changes.
FAQ
Is Gemini 3.1 Pro publicly confirmed as an MoE model?
Gemini 3.1 Pro's official model card says that it is based on Gemini 3 Pro but does not disclose enough detail to support a specific MoE configuration. Architecture inference should not be presented as fact without public evidence. It is also no longer the current flagship; that role has moved to Gemini 4 Argon.
Can Veo 3.1 generate 90 seconds in one request?
The official product page confirms scene extension and 4K output, while its public comparisons mainly use six-to-eight-second clips. This revision does not list 90 seconds as a universal single-generation limit. Check the exact entry point you use.
Can the five model families form an automated pipeline?
They can be orchestrated through APIs and product surfaces, but every connection still requires validation of format, semantic loss, latency, permissions, and recovery. A shared brand is not a completed system integration.
When should this article be checked again?
When Google changes model IDs, preview status, pricing, access paths, context or output limits, or robotics safety restrictions. Model cards can also change without a model-name change.
References
- Google DeepMind, Gemini 4 Argon model card
- Google DeepMind, Gemini 3.1 Pro model card
- Google AI for Developers, Gemini API deprecations
- Google DeepMind, Veo 3.1
- Google AI for Developers, Veo 3.1 API lifecycle
- Google DeepMind, Lyria 3.5
- Google DeepMind, Lyria 3.5 model card
- Google AI for Developers, Music generation models
- Google DeepMind, Genie 4 announcement
- Google DeepMind, Genie 3 announcement
- Google DeepMind, Gemini Robotics 2
- Google DeepMind, Gemini Robotics ER 2 model card