Google and DeepMind AI Models in 2026: Gemini 3.8, Gemma 4, Veo, Nano Banana, Robotics, Genie, and Alpha Systems

[!NOTE] Current as of September 22, 2026.
An overview of the Google and Google DeepMind ecosystem, technical specifications, and model availability. Every technical claim in this guide is based on official Google, Google DeepMind, Google AI for Developers, and Vertex AI documentation.

This guide provides a systematic overview of the most important active, publicly introduced, and announced Google and Google DeepMind AI models and research systems in 2026.

In 2026, artificial intelligence is no longer limited to one isolated language model that answers text questions. The Google and Google DeepMind technology ecosystem has branched into a specialized network of tools: production multimodal models for agentic work and coding, systems for voice interaction and video generation, open Gemma models, embodied intelligence in robotics, and fundamental scientific breakthroughs in genomics and mathematics.

Quick status overview as of September 22, 2026

To make it easy to distinguish production API models from laboratory experiments, each system in this article has a specific status label:

• [GA / Stable]: Generally available production model with official API support and defined pricing.
• [Preview / Experimental]: Publicly available for testing and early development before full commercial availability.
• [Research / Limited Access]: Research system demonstrated in scientific papers or a model with access limited to verified partners.
• [Confirmed in Development]: Project officially confirmed to be in active development or pre-training.
• [Operational / Production]: Model Google uses in operational and production services or publicly distributes through platform data channels, but not necessarily a standard generative API model with token-based pricing.
• [Legacy / Previous Generation]: An older model generation being gradually replaced by newer releases.

Key active model lines in September 2026

• Main production frontier API: Gemini 3.8 Flash, Gemini 3.8 Live, with Gemini 3.1 Pro in Preview.
• Specialized and agentic capabilities: Gemini 3.8 Flash Cyber, Gemini Computer Use, Deep Research.
• Multimodal generation (images, video, audio): Nano Banana 2 (and Pro / Lite), Veo 3.1, Lyria 3.5, Gemini Omni 1.1 Flash.
• Open weights: Gemma 4 family (E2B, E4B, 12B, 26B A4B MoE, 31B).
• Embodied AI and simulations: Gemini Robotics 2, Gemini Robotics-ER 2, Genie 3, SIMA 2.
• Science, biology, and climate: AlphaFold 3, AlphaGenome and AlphaGenome Atlas, WeatherNext 3, AlphaEarth Foundations, AlphaEvolve.
• In development: Gemini 3.5 Pro (Coming soon), Gemini 4 (Pre-training phase).

01What in this ecosystem is not a classic foundation model?

In public discussions, the word “model” is often used for everything that comes out of an AI lab. For a technically accurate overview, it is important to distinguish foundation models from platforms, agents, and tools:

• Google Antigravity: Not a standalone model, but an agent-first development platform and agent harness that integrates Gemini models, local tools, and a terminal.
• Gemini Computer Use: Not a separate foundation model, but a tool or capability for visual interface control used by supported Gemini models, primarily Gemini 3.8 Flash.
• Gemini Deep Research: Not a foundation language model, but an autonomous research agent that orchestrates search, extraction, and synthesis.
• CodeMender: An AI security agent for automated code analysis and security patch generation.
• AI Co-Scientist: A multi-agent research system that simulates scientific debate and hypothesis generation.
• AlphaEvolve: An evolutionary agent system for discovering new computational algorithms.
• AlphaFold Server: A web service and research platform that provides access to the AlphaFold 3 model.
• Gemma Scope 2: Not a language model, but an interpretability toolkit for analyzing internal activations and neural network safety.

02Gemini: the main frontier model family

The Gemini model family is the primary API choice for software engineers and enterprise systems through Google AI Studio and Google Cloud Vertex AI.

                  [GEMINI FRONTIER ARCHITECTURE]
                                │
       ┌────────────────────────┼────────────────────────┐
       ▼                        ▼                        ▼
 [Working Flash Models]  [Advanced Reasoning]   [Long Context / Pro]
  Gemini 3.8 Flash        Gemini 3.1 Deep Think  Gemini 3.1 Pro
  Gemini 3.8 Live         Gemini 3.8 Live Ext.   Gemini 3.5 Pro (Coming soon)
  Gemini 3.5 Flash-Lite   AlphaEvolve (Agent)    Gemini 4 (Pre-training)

Gemini 3.8 Flash [GA / Stable]

Official API identifier: gemini-3.8-flash. Released on September 2, 2026, as the leading general-purpose model in the Flash series.

• Official purpose: Long-horizon software engineering, autonomous agents, complex business workflows, multi-step reasoning, and multimodal processing of large contexts.
• Input context window: 1,048,576 tokens (1M).
• Maximum output limit: 65,536 tokens (64K).
• Supported input formats: Text, images, video, audio, and PDF documents.
• Output format: Text.
• Thinking levels: Supports low, medium, and high. The minimal level is not supported.
• Supported tools and API features: Caching, Code Execution, File Search, Function Calling, Google Maps Grounding, Google Search Grounding, Structured Outputs (JSON Schema), URL Context, and Computer Use support.
• Official prices (Google AI API, September 2026):
  • Introductory rate, valid through December 31, 2026: $0.75 per million input tokens; $3.75 per million output tokens.
  • Announced standard rate, starting January 1, 2027: $1.50 per million input tokens; $7.50 per million output tokens.

Gemini 3.8 Flash Cyber [Research / Limited Access]

A specialized variant of Gemini 3.8 Flash for defensive cybersecurity.

• Official capabilities: Autonomous discovery of security flaws in source code, automated patch generation and testing, and analysis of complex repositories across many programming languages.
• Access: The model is not part of the public API. Access is limited to verified security professionals through Google's defensive program, Fairwind.
• Scope: It is primarily used for defensive purposes. Scenarios such as malware analysis and reverse engineering are practical applications within Fairwind, not a publicly specified training corpus.

Gemini 3.8 Live [GA / Stable]

Released on September 15, 2026, as an advanced solution for low-latency, two-way voice and multimodal interaction in real time.

• Architecture: Native audio-to-audio processing without separate intermediate speech-to-text and text-to-speech steps.
• Input context: 131,072 tokens; maximum output: 65,536 tokens.
• Supported inputs: Text, images, audio, and streaming video; output includes text and native audio.
• Capabilities: Asynchronous function calling, live visual context from a camera or screen, reasoning during conversation, and support for more than 97 languages.

Gemini 3.8 Live Extended Thinking [GA / Stable]

An extended version of the Live model that can handle more complex tasks during a voice session.

• Purpose: Background reasoning, planning more complex steps, and asynchronous tool execution while the live conversation continues without interruption.
• Configuration: Uses the thinkingLevel parameter (low, medium, or high) to balance analysis depth and spoken response speed.

Gemini 3.5 Pro [Confirmed in Development]

A model that Google and DeepMind officially list as Coming Soon. It is undergoing internal and partner testing. As of September 22, 2026, it is not publicly available as a general production API, and official data about price, exact context window, and performance measurements have not been released.

Gemini 3.1 Pro [Preview]

An advanced Gemini Pro model currently available in Preview, designed for demanding analytical, multimodal, and programming tasks. Official API identifier: gemini-3.1-pro-preview.

• Context limits: 1,048,576 input tokens; 65,536 output tokens.
• Practical use: Its large context makes it suitable for analyzing entire code repositories, long video and audio recordings, and extensive technical documentation without artificially splitting them into sections.

Gemini 3.1 Deep Think [Preview / Research]

An advanced reasoning mode designed for complex problems in mathematics, physics, chemistry, engineering, and algorithm design that require rigorous verification of each logical step.

Context capacity of previous generations: Gemini 2.5 series [Legacy / Stable]

Gemini 2.5 Pro and Gemini 2.5 Flash have a standard input window of 1,048,576 tokens (about one million, not multiple millions). The older Gemini 1.5 Pro offered up to 2,097,152 tokens in an experimental phase, but belongs to a previous generation.

Gemini 4 [Confirmed in Development]

The next major milestone in the development of Google's frontier models.

• Officially confirmed: Google has publicly confirmed it has started its most ambitious pre-training cycle yet for Gemini 4.
• Public specifications as of September 22, 2026: Google has not published an official model card, API identifier, context size, output limit, price, benchmark results, or final architecture.

[!WARNING] Unconfirmed rumors and unofficial claims: Figures circulating in forums, such as an alleged 10 million token context, a 256K output limit, persistent cross-session memory, or a standalone execution environment, are unconfirmed rumors [UNVERIFIED] and are not part of official Google documentation.

03Gemini Live, Speech, and the Audio ecosystem

Real-time speech processing has moved to native models that replace older multi-step transcription and speech synthesis pipelines.

04Gemini Omni: strategic direction and current status

Google describes Gemini Omni as a step toward a long-term vision of a model that can create and edit different media types from different kinds of input, starting with video.

Long-term vision (strategic goal):       Current state (Gemini Omni 1.1 Flash):
[Text / Image / Audio / Video]            [Text / Image / Audio / Video]
               │                                         │
               ▼                                         ▼
      [ Universal Omni ]                     [ Omni 1.1 Flash Engine ]
               │                                         │
               ▼                                         ▼
[Text / Image / Audio / Video]                [High-res video with native audio]

Gemini Omni 1.1 Flash [Preview / Experimental]

The current working implementation of the Omni concept.

• Input modalities: Text, images, audio, and video.
• Current output: High-resolution video with integrated audio.
• Confirmed capabilities:
  • Generate video from text and image descriptions.
  • Conversational editing of existing video material.
  • Extend the duration of video shots.
  • Interpolate intermediate frames (keyframe interpolation).
  • Upscale video resolution.

05Image generation and editing: Nano Banana / Gemini Image

The image creation and semantic editing model line is part of the Gemini Image ecosystem, known as Nano Banana. These models are available through the Gemini API, Google AI Studio, and Vertex AI.

Nano Banana Pro (Gemini Pro Image) [GA / Stable]

The leading model for high-fidelity image generation.

• Precise text rendering: Reliably renders labels, typography, symbols, and longer text inside an image.
• Complex spatial layouts: Follows detailed instructions about object placement in a scene.
• Semantic editing: Supports inpainting and outpainting while preserving existing style and lighting.

Nano Banana 2 (Gemini Flash Image) [GA / Stable]

A model optimized for interactive applications where speed matters. Supports 0.5K, 1K, 2K, and 4K resolutions, aspect ratio control, integrated reasoning (Thinking), Search Grounding, and the use of video as multimodal context for image generation.

Nano Banana 2 Lite [GA / Stable]

The least expensive and fastest variant. It is a cloud/API model optimized for frequent use and fast interactive editing. Note: “local edits” means localized changes to a specific part of an image, not local on-device inference on a phone.

Nano Banana (Gemini 2.5 Flash Image) [Legacy]

The earlier generation, gemini-2.5-flash-image, for which Google recommends migrating to the Nano Banana 2 series.

06Video generation: Veo 3.1

The Veo model family is designed for professional motion image creation with high temporal coherence.

[Text prompt or starting image]
                 │
                 ▼
          [ VEO 3.1 MODEL ]
                 │
      ┌──────────┴──────────┐
      ▼                     ▼
[Up to 4K video]       [Native synchronized audio]
- Camera and motion control  - Ambient sound and effects
- Temporal coherence        - Improved audio-video alignment
- Add or remove objects     - Dialogue (active development area)

• Veo 3.1 [Preview, Gemini API]: Official API endpoint: veo-3.1-generate-preview, with the fast variant veo-3.1-fast-generate-preview. The main model for Text-to-Video and Image-to-Video. Veo 3.1 supports 1080p and 4K output for supported generations and durations, for example up to 8 seconds. It has built-in synchronized audio, including ambience, sound effects, and dialogue with improved timing alignment. It offers director-level camera control, scene extension, first and last frame control, outpainting, and selective addition or removal of objects from a shot.
• Veo 3.1 Lite [Preview, Gemini API]: Official API endpoint: veo-3.1-lite-generate-preview. A faster version optimized for quick concept checks and previsualization.
• Veo 3 and Veo 2 [Legacy]: Previous generations that established standards for temporal consistency.

07Music and audio generation: the Lyria ecosystem

Developed in collaboration with music creators, Lyria generates instrumental parts and songs.

• Lyria 3.5 [GA / Stable]: The main current generation, stable since September 3, 2026. It generates complete stereo music tracks from text or image prompts, with support for lyrics.
• Lyria 3 Pro [Preview / Legacy]: A previous release for studio testing. Note: multitrack instrument exports in demonstrations were part of application workflows, not a native model feature.
• Lyria 3 Clip [Preview]: Official API endpoint: lyria-3-clip-preview. Generates short music clips and loops up to about 30 seconds.
• Lyria RealTime [Preview]: An interactive model that dynamically changes music in real time in response to external events, such as action levels in a game.
• Lyria 2 [Legacy]: An older version. Audio labeling uses SynthID watermarking, part of Google's broader safety standard.

08Gemma: the open-weight model family

For engineers and companies that want to host models on their own servers without sending data to external services, Google develops the open-weight Gemma family.

                    [ GEMMA 4 FAMILY ]
                             │
       ┌─────────────────────┼─────────────────────┐
       ▼                     ▼                     ▼
 [Edge & Mobile]      [Server Dense]        [MoE Scaling]
 Gemma 4 E2B (Edge)    Gemma 4 12B           Gemma 4 26B A4B (MoE)
 Gemma 4 E4B (Edge)    Gemma 4 31B

Gemma 4 series [GA / Open-Weight]

• Gemma 4 E2B and E4B: 128K token context window. Modalities: Text, Image, Audio → Text. Optimized for energy-efficient operation on mobile devices and edge devices.
• Gemma 4 12B: 256K token context window. Modalities: Text, Image, Audio → Text. A general-purpose model for locally hosted systems, multimodal analysis, and coding.
• Gemma 4 26B A4B (MoE): 256K token context window. Modalities: Text, Image → Text. A Mixture-of-Experts architecture where 4 billion of 26 billion parameters are activated per token, reducing compute during inference.
• Gemma 4 31B: 256K token context window. Modalities: Text, Image → Text. The largest dense open-weight model in this line, designed for complex reasoning and coding.

Broader family of specialized Gemma models

1. FunctionGemma [GA]: Tuned to recognize function formats and call external tools.
2. EmbeddingGemma [GA]: A compact model for creating semantic vectors locally.
3. DiffusionGemma [Experimental / Research]: An experimental model exploring diffusion mechanisms for text generation instead of autoregressive layers.
4. MedGemma and MedGemma 1.5 4B [Research / Open-Weight]: Models optimized for understanding medical text and analyzing radiology images. They are not intended to make clinical diagnoses on their own without verification by a qualified physician.
5. TranslateGemma [GA]: Specialized for translating text into 55 world languages.
6. T5Gemma [GA]: An encoder-decoder architecture suitable for data extraction and search.
7. ShieldGemma and ShieldGemma 2 [GA]: Safety classifiers for filtering text and multimodal content.
8. CodeGemma [GA]: Specialized in code analysis and generation.
9. PaliGemma and PaliGemma 2 [GA]: Vision-language models (VLMs) for object detection and visual question answering.
10. RecurrentGemma [Research]: A hybrid architecture based on the Griffin mechanism, with fixed memory consumption for long sequences.
11. DolphinGemma [Research]: A scientific model created with biologists for acoustic analysis and research into dolphin communication.
12. Gemma 3 270M [GA]: An exceptionally small model for very lightweight classification tasks.
13. DataGemma [Research]: Connects models to public statistical data through the Data Commons platform.
14. TxGemma [Research / Open-Weight]: Open models for early-stage drug discovery, covered in more detail in the biology section.

09Robotics: Gemini Robotics (Embodied AI)

Applying vision-language models to control physical robots uses a Vision-Language-Action (VLA) architecture.

[Camera and sensors] ──→ [Gemini Robotics-ER 2] ──→ [Multi-step plan]
                             │
                             ▼
               [Gemini Robotics On-Device 2]
                             │
                             ▼
                 [Motors and manipulators]

• Gemini Robotics 2 [Research / Limited Access]: A native VLA model for directly controlling robot bodies and manipulators in the physical world.
• Gemini Robotics-ER 2 [Research]: An embodied-reasoning VLM with a 128K context window and 64K output. It analyzes text, images, video, and audio to perform spatial, temporal, and physical reasoning and long-term task planning.
• Gemini Robotics On-Device 2 [Research]: A model for local execution on a robot without an internet connection. In research, it adapted to new robot configurations with fewer than 200 demonstrations and a few hours of training data.

10World models and simulations: Genie 3 and SIMA 2

Research focused on understanding spatial and physical rules through simulation:

Genie 3 [Research]

A general World Model that generates an interactive 3D environment in real time from a text description or a single image, at about 20 to 24 frames per second in 720p. It enables real-time navigation and dynamic events in generated spaces, with temporal consistency lasting several minutes. It is a neural world model, not a conventional game engine.

SIMA 2 (Scalable Instructable Multiworld Agent) [Research]

A general AI agent that uses Gemini as its reasoning core. It sees the world only through screen pixels and controls standard inputs such as a keyboard and mouse, without access to the game's internal code. It can follow multimodal instructions in completely unfamiliar virtual environments and has been demonstrated in integration with Genie 3 simulations.

11Multimodal vector space: Gemini Embedding 2

For semantic search systems and RAG (Retrieval-Augmented Generation) architectures:

• Gemini Embedding 2 [GA / Stable]: Official API identifier: gemini-embedding-2.
  • Input: Text, images, video, audio, and PDF documents.
  • Input limit: 8,192 tokens.
  • Output vector dimensions: Flexible, from 128 to 3,072 dimensions.
  • Purpose: Semantic search, multimodal retrieval, clustering, and cross-media comparison in a unified vector space.
• Gemini Embedding 001 [GA / Stable]: A standard text model for text knowledge bases.
• EmbeddingGemma [GA / Open-Weight]: An open model for generating vectors on local infrastructure.

12Computer Use and agent systems

The shift from passive answers to active task execution in a software environment:

[User task] ──→ [Gemini 3.8 Flash + Computer Use Tool]
                                │
       ┌────────────────────────┼────────────────────────┐
       ▼                        ▼                        ▼
[Screen capture]          [Reasoning]          [Mouse/keyboard action]
- Identify elements       - Plan next step     - Click, type, move pointer
- Read tables and forms

Gemini Computer Use capability [Preview]

A feature or tool that lets supported models, with Gemini 3.8 Flash recommended, control graphical interfaces in browser, mobile, and desktop environments. It works in a loop: screenshot → planning → action → verification of the new state. It has built-in defenses against prompt injection and asks for user confirmation for sensitive actions. Previous endpoint: gemini-2.5-computer-use-preview-10-2025.

Gemini Deep Research and Deep Research Max [Preview / Agent]

Autonomous research agents, with API endpoints deep-research-preview-04-2026 and deep-research-max-preview-04-2026. They have a 1M context and 65K output. They search the web and local files (File Search), use URL Context, MCP tools, and code execution to cross-analyze sources and generate structured research reports with citations and charts.

CodeMender [Research / Limited Access, Security Agent]

An AI security agent that combines static analysis, dynamic testing, fuzzing, differential testing, and SMT solvers to automatically find vulnerabilities and generate patches that undergo engineering review before integration.

AI Co-Scientist [Research, Multi-Agent System]

A system based on Gemini models that coordinates collaboration among specialized agents. Agents generate hypotheses, critique one another's proposals, and rank and synthesize research ideas through a scientific debate format.

13Biology, medicine, and genomics: Alpha systems

DeepMind's life science systems map biological structures and genetic code.

Biological level:       System:                         Prediction scope:
[3D complexes]          AlphaFold 3                     Proteins, DNA, RNA, ligands
[Point mutations]       AlphaMissense                   71 million missense variants
[Entire genome]         AlphaGenome & Atlas             ~9 billion nucleotide variants
[Therapeutic molecules] TxGemma (2B, 9B, 27B)           Small molecules, antibodies, proteins
[Clinical text and image] MedGemma                      Medical NLP and radiology

AlphaFold 3 [GA / Scientific Access]

A model that predicts the three-dimensional structures and interactions of almost all biological molecules: proteins, DNA, RNA, and ligands (small molecules important to drug development).

AlphaGenome and AlphaGenome Atlas [Research]

One of DeepMind's major scientific breakthroughs. AlphaGenome analyzes regulatory DNA sequences and predicts how genomic changes affect cell function. AlphaGenome Atlas contains predictions for approximately 9 billion single-nucleotide variants, practically covering every possible single-base mutation in the human genome. Google DeepMind has released a research implementation and pretrained AlphaGenome weights, available to researchers through Kaggle and Hugging Face subject to special non-commercial licensing terms.

AlphaMissense [Research]

Classified approximately 71 million possible human missense mutations, placing about 89% into likely benign or likely pathogenic categories and providing a resource for rare disease research.

TxGemma [Research / Open-Weight]

A family of open models in 2B, 9B, and 27B sizes, trained on about 7 million examples for therapeutic development and evaluated on more than 60 tasks involving small molecules, proteins, nucleic acids, and cell lines.

14Weather forecasting, climate, and Earth

AI weather models generate forecasts much faster than traditional numerical methods and achieve competitive or better results on specific meteorological benchmarks.

• WeatherNext 3 [Operational / Production]: Uses direct satellite data for global weather forecasts updated every hour. It provides station-targeted temperature and dewpoint predictions at about 0.05° (~5 km) resolution, while standard gridded surface variables such as temperature and wind are available at about 0.1° (~10 km), depending on the specific weather product and output.
• GenCast [Research]: A probabilistic diffusion model that creates forecast ensembles up to 15 days ahead, optimized for early detection of extreme weather events.
• GraphCast [Research]: A graph neural network that produces a global 10-day forecast in less than one minute on a single TPU processor.
• NeuralGCM [Research]: A hybrid model combining numerical fluid dynamics with machine learning for long-term climate projections.
• AlphaEarth Foundations [Research]: A geospatial foundation model for Earth observation. It maps gigabytes of satellite imagery into a shared 64-dimensional embedding space, covering over 1.4 trillion spatial points annually, to track changes in vegetation, agriculture, and waterways.

15Mathematics, algorithms, and fundamental science

• AlphaEvolve [Research]: An evolutionary agent system powered by Gemini models for discovering new algorithms. In Google's production data centers, it found a new heuristic for Borg job scheduling that freed approximately 0.7% of Google's total computing resources.
• AlphaGeometry 2 and AlphaProof [Research]: At the International Mathematical Olympiad (IMO 2024), these systems reached a silver medal equivalent, scoring 28/42 points. AlphaProof formalizes mathematical statements in Lean and independently assembles formally proven steps. Later, at IMO 2025, the newer Gemini Deep Think reached gold medal level with 35/42.
• FunSearch [Research]: A system that pairs a language model with an automatic evaluator to find new mathematical solutions. It has been applied successfully to the cap set problem and online bin-packing heuristics.

16History and archaeology: Aeneas

• Aeneas [Research]: An AI system for studying ancient written heritage, trained on a collection of more than 176,000 Latin inscriptions.
  • Restoration accuracy (Top 20): about 73% for damage up to 10 characters, or about 58% when the missing text length is not known in advance.
  • Geographic origin: about 72% accuracy in locating inscriptions across 62 Roman provinces.
  • Dating: estimates within about 13 years of the historical consensus.
  • Key innovation: contextualization and parallel search, linking damaged inscriptions to related historical texts.

17Historical milestone systems in games (RL)

DeepMind's early breakthroughs in reinforcement learning:

• AlphaGo, AlphaGo Zero, and AlphaZero [Historical / Research]: Mastered Go, chess, and shogi through self-play.
• MuZero [Historical / Research]: Learns strategy and rules without prior knowledge. In practice, it has been used in YouTube live infrastructure and saved about 4% in bitrate during VP9 video compression.
• DreamerV3 and Dreamer 4 [Historical / Research]: Learn through an internal world model. DreamerV3 was the first agent to find a diamond from scratch in Minecraft without human demonstrations.

18Large comparison matrix of models and systems

The following table brings together key technical characteristics for the major Google and Google DeepMind systems in 2026:

Google DeepMind has released a research implementation and pretrained AlphaGenome weights through Kaggle and Hugging Face under special non-commercial licensing terms.

19Conclusion

In 2026, Google and Google DeepMind are not developing one monolithic model. They are building a hierarchy of systems optimized for different computing needs:

• The Flash family is Google's main line optimized for a combination of speed, cost, and production agent workloads.
• The Live and Omni lines shift interaction toward native speech and video processing without intermediate steps.
• The Gemma series gives the developer community open weights for local privacy and independent control.
• Alpha systems apply specialized AI methods to biology, genomics, mathematics, algorithms, geospatial data, and fundamental science.

Google has officially confirmed development of Gemini 4, but its final architecture and how it will combine existing multimodal, voice, agentic, and robotics capabilities have not yet been specified publicly.