Skip to content
AI Department Store

Google Cloud

Don't bet on the winning model.Pick the right store.

No enterprise wants to bet its AI strategy on one model, one vendor or one generation of technology. Everyone wants one trusted place that carries the best option for every need.

ROOFTOPPeopleFLOOR 5AgentsFLOOR 4ModelsFLOOR 3Data + ToolsFLOOR 2Build + RuntimeFLOOR 1Govern + OptimizeFOUNDATIONAI HypercomputerIDENTITY • SECURITY • GOVERNANCE • OPEN STANDARDS

01 · The customer problem

The bet nobody can make.

Every enterprise is being asked to place the same bet. Which model, which vendor, which architecture will still be right in three years?

  1. Which model will be best in twelve months?

    Nobody knows. Not us, not them. The leaderboard has changed hands repeatedly and will again.

  2. Which vendor will lead in three years?

    Every provider has had a great season. None has had every season.

  3. Which architecture survives the next generation?

    Only the one that does not depend on the answer to the first two questions.

What enterprises actually want

  • One trusted place.
  • The best option for every need.
  • The freedom to change their mind without rebuilding.

02 · The analogy

Think of the best department storein your city.

You don’t go there because it makes everything. You go because it chooses well, and everything works under one roof.

  1. ChoiceEvery serious brand, side by side on the same floor.
  2. CurationThe buyers have already done the sorting. Not one of everything; the right things.
  3. One buildingOne entrance, one card, one returns desk, one loyalty programme.
  4. One standardThe same service, security and guarantee whichever brand you pick.

Equals

Optionality without chaos

The insight

Google Cloud is the department store of enterprise AI. Gemini is our house brand and it sits on the front rack. But the same shelves carry Claude, Grok, Mistral, DeepSeek, Gemma and hundreds more, and every one of them runs on the same building: the same security, the same data, the same operations, the same bill.

The question for the board is not which model to bet on. It is which store to shop in.

03 · Why optionality matters

Switch the model.Not the architecture.

Optionality is the hero of this story. You don’t know which model will win tomorrow. In a department store, you don’t have to.

One use case. Multiple choices.

Use case

Deep reading across long documents and many sources. Frontier reasoning earns its cost here.

Models that could do the job

Switched 0 times. Everything beneath the model stayed put.

The product on the shelf changes. The building doesn’t. That is what optionality means once it leaves the slide and enters the architecture.

Model names as documented on Agent Platform, 5 September 2026. Preview models are pre-GA.

Application / agentResearch agent
Gemini 3.1 ProFrontier · 1M-token context · Preview
Common Google Cloud platformUnchanged
Data
Security
Governance
Operations
Infrastructure

04 · Why this is different

Betting on one house,or planning for change.

A single-brand store asks you to believe one house wins every category forever. A department store assumes innovation stays fragmented, and gives you whichever brand wins. The difference is not which model is best this quarter; it is what each strategy assumes about the future.

The single-brand store

Assumes one provider will win every category, every year.

  • Your whole wardrobe follows one house’s taste, pricing and roadmap.
  • A bad season for them is a bad season for you.
  • Changing your mind means moving house.

The department store

Assumes innovation will stay fragmented, and plans for it.

  • Whichever brand wins a category is already on the shelf.
  • You change the product, not the building.
  • Security, data and operations are the store’s job, not the brand’s.

Google’s differentiation

Not “our model is always the best.”

The best model for the job today. A different one tomorrow if the market moves. One enterprise platform throughout.

05 · The whole idea

The whole idea, in ninety seconds.

Simple, visual, repeatable. Read it before anyone explains the technology; the rest of this page only proves it.

  1. Every enterprise is being asked to make the same bet: pick the model, pick the vendor, pick the architecture that will still be right in three years. Nobody can make that bet honestly.

  2. So think of the best department store in your city. You don’t shop there because it makes everything. You shop there because it chooses well, and everything works under one roof: one entrance, one card, one returns desk, one standard of service whichever brand you pick.

  3. Google Cloud is the department store of enterprise AI. Gemini is our house brand, and it’s on the front rack. But the same shelves carry Claude, Grok, Mistral, DeepSeek, Gemma and hundreds more. Same checkout, same security, same audit trail, same contract.

  4. Underneath is one platform. Your data grounds every answer in your business. Security and governance wrap the whole building, so adding a model never means adding a risk review. And the doors open both ways: bring what you’ve built, take out what you build here.

  5. The point is simple. You shouldn’t have to predict the winner. You should be able to choose the best capability for every job today, change it tomorrow when the market moves, and never rebuild the store.

Share them as they are. They are the whole argument.

06 · Map the store to the platform

Every part of the storehas a name on the platform.

Walk the building from the front door down. Each part of the store is one part of Google Cloud’s AI platform.

In the storeOn Google CloudWhat it means for you
The store itselfGoogle Cloud AI platform: Gemini Enterprise Agent PlatformOne operating environment across build, scale, govern and optimise. The building everything else sits in.
The front door and the conciergeGemini EnterpriseWhere employees enter, ask, find agents and get work done, with their existing permissions.
The brands and departmentsGoogle models, partner models, open models, specialised capabilitiesGemini on the front rack; Claude, Grok, Mistral, DeepSeek, Gemma and specialists for image, video, speech and code alongside.
The breadth of the shelfModel GardenMore than 200 models under one roof, in four tiers: frontier, workhorse, efficient, specialist.
The customer’s own measurementsData and grounding: BigQuery, enterprise search, RAG, connectorsThe context that decides what is relevant. Your data makes every answer yours, and it is never used to train public models.
The finished outfitAgents and applications: Agent Garden, Agent Studio, ADK, Agent RuntimeWhat the customer actually consumes: an agent that does a job from start to finish, assembled from models, tools and data.
Security, standards and quality controlAgent Identity, Agent Gateway, Model Armor, evaluation, observabilityCommon standards across the entire store. Add a brand and it inherits the controls.
The building, logistics and supply chainAI Hypercomputer: TPUs, GPUs, networking, storageEverything customers don’t see but depend on: performance, scale and economics.

Walk the floors

Seven levels. One structural column. Click a floor.

  1. The model floor: house brand, premium brands, independent brands

    Model Garden is one place to discover more than 200 models. Google's own line runs from Gemini 3.1 Pro (frontier) through Gemini 3.8 Flash (the workhorse) to Gemini 3.5 Flash-Lite (efficient), with Veo, Imagen and Lyria as specialists. Partner brands such as Anthropic Claude, xAI Grok and Mistral AI sit on the same shelf as managed APIs. Open models from Google, DeepSeek, Zhipu, Moonshot and Alibaba are managed or self-deployed. The store carries its own flagship label and does not force you to buy it.

    Model GardenGemini 3.1 Pro · 3.8 Flash · 3.5 Flash-LiteClaude Fable 5.1 · Opus 5 · Sonnet 5 · Haiku 4.5Grok 4.6 · Codestral 2Gemma 4 · DeepSeek-V3.2 · GLM 5.2 · Kimi K2 · Qwen3-Next

07 · Customer use cases

Follow one AI ideathrough the store.

An agent that researches a customer, analyses internal data, prepares a recommendation and initiates follow-up actions. The same walk works in any industry: the products change, the building does not.

ROOFTOPPeopleFLOOR 5AgentsFLOOR 4ModelsFLOOR 3Data + ToolsFLOOR 2Build + RuntimeFLOOR 1Govern + OptimizeFOUNDATIONAI HypercomputerIDENTITY • SECURITY • GOVERNANCE • OPEN STANDARDS

Step 1 of 10

Enter

Gemini Enterprise

Application / agentCustomer recommendation agent
Gemini 3.8 FlashWorkhorse tier · chosen on the model floor
Common Google Cloud platformUnchanged
Data
Security
Governance
Operations
Infrastructure

Proof 01 · Floor 4

Shop by outcome.Not by brand.

Model Garden is one place to discover more than 200 models. The shelf below is this generation only, sorted the way a buyer thinks: by the job, not the logo.

How to read the shelf

Four tiers run across every shelf. Premium is a tier, not a brand: each shelf carries a frontier model, a workhorse and an efficient option. Featured models were released within the last twelve months. Older families stay on the shelf but are not shown here.

  1. Frontier

    The hardest problems.

    Deepest reasoning, long-horizon agents, highest cost per token. Use where the answer is worth it.

  2. Workhorse

    The everyday default.

    Near-frontier quality at a cost profile that runs agents at scale. Most production work lives here.

  3. Efficient

    Volume and latency.

    Classification, routing, extraction, real-time service. Fast, cheap, good enough by design.

  4. Specialist

    One job, done well.

    Image, video, music, speech, code, documents, embeddings. Built for a task rather than every task.

Google

The flagship in-house label

Gemini and Google’s specialised models. Deeply integrated with the rest of the store, and never the only thing on the shelf.

Access: Managed APIs on Agent Platform. Gemma 4 is also available as an open model.

FrontierThe hardest problems.
  • Gemini 3.1 ProGoogle· Feb 2026Preview

    Most advanced reasoning model; 1M-token context; text, audio, image, video, PDF and whole code repositories.

WorkhorseThe everyday default.
  • Gemini 3.8 FlashGoogle· Sep 2026

    Most intelligent workhorse model; software engineering, agentic tasks, multi-step reasoning; often approaches frontier performance at lower cost.

EfficientVolume and latency.
  • Gemini 3.5 Flash-LiteGoogle· Jul 2026

    Cost-effective line for simple coding, precise document understanding and lightweight agents; built for high-throughput classification and extraction.

SpecialistOne job, done well.
  • Gemini 3.1 Flash ImageGoogle· May 2026

    Image understanding and generation at a balance of price and performance.

  • Gemini 3 Pro ImageGoogle· Nov 2025

    Text-to-image for the highest-fidelity creative work.

  • Gemini Omni 1.1 FlashGoogle· Aug 2026Preview

    Video, image and text in one model, with video output alongside text.

  • Veo 3.1Google· Oct 2025

    Text-to-video and image-to-video.

  • Lyria 3Google· Feb 2026

    Music generation.

  • Gemini 3.5 Transcribe · Live TranslateGoogle· Jul 2026

    Speech to text and live translation in the Gemini 3.5 line.

  • Gemini Embedding 2Google· Apr 2026

    Natively multimodal embeddings for search, retrieval and grounding.

Partner models

Premium brands carried inside the store

Frontier models from other providers, sold as managed APIs inside the same environment, governed by the same platform, billed on the same account.

Access: Managed APIs (model as a service). No infrastructure to run.

FrontierThe hardest problems.
  • Claude Fable 5.1Anthropic· Sep 2026

    Anthropic’s newest top-tier model: autonomous knowledge work and coding; long-running, complex and asynchronous tasks.

  • Claude Opus 5Anthropic· Jul 2026

    Most advanced Opus model: long-running agents, ambitious coding, deep professional and financial analysis, computer use.

  • Grok 4.6xAI· Aug 2026Preview

    xAI’s most capable model for coding, agentic tasks and knowledge work.

WorkhorseThe everyday default.
  • Claude Sonnet 5Anthropic· Jun 2026

    Most capable Sonnet yet; lead agent or sub-agent in production pipelines with the cost profile to run high-volume agentic work.

  • Grok 4.20xAI· Mar 2026Preview

    Reasoning and non-reasoning variants; document understanding and long-horizon tool calling.

EfficientVolume and latency.
  • Claude Haiku 4.5Anthropic· Oct 2025

    Anthropic’s current small model: near-frontier performance at the speed and cost for service agents, sub-agents and high-volume experiences.

  • Grok 4.1 FastxAI· Nov 2025Preview

    xAI’s most cost-effective model; search, summarisation and categorisation at volume.

SpecialistOne job, done well.
  • Codestral 2Mistral AI· Apr 2026

    Code generation and fill-in-the-middle completion.

Open models

Independent brands and specialist products

Open-weights choices from Google and the wider community. Managed as a service, or, for supported models, self-deployed into your own environment with your own weights.

Access: Managed APIs, or one-click self-deployment for supported models.

FrontierThe hardest problems.
  • DeepSeek-V3.2DeepSeek· Dec 2025

    Computational efficiency with strong reasoning and agent performance; the reasoning tier of the open shelf.

  • Kimi K2 ThinkingMoonshot AI· Nov 2025

    Open thinking-agent model that reasons step by step and uses tools.

  • Qwen3-Next-80B ThinkingAlibaba· Sep 2025

    Complex problem-solving and deep reasoning in the Qwen3-Next family.

WorkhorseThe everyday default.
  • GLM 5.2Zhipu AI· Jun 2026

    Long-horizon agentic and coding tasks with a 1M-token context window; the newest open workhorse on the shelf.

EfficientVolume and latency.
  • Gemma 4 26BGoogle· Apr 2026

    Google’s open multimodal model; a house label you can take with you.

  • Qwen3-Next-80B InstructAlibaba· Sep 2025

    Instruction-following at a small active-parameter cost.

SpecialistOne job, done well.
  • DeepSeek-OCRDeepSeek· Oct 2025

    Optical character recognition for complex documents.

  • MiniMax M2MiniMax· Oct 2025

    Agentic and code tasks: planning and executing complex tool calls.

The shelf at a glance

Names, tiers and release months verified against Google Cloud documentation and provider announcements on 5 September 2026.

TierGooglePartner modelsOpen models
FrontierThe hardest problems.Gemini 3.1 Pro (Preview)Claude Fable 5.1 · Claude Opus 5 · Grok 4.6 (Preview)DeepSeek-V3.2 · Kimi K2 Thinking · Qwen3-Next-80B Thinking
WorkhorseThe everyday default.Gemini 3.8 FlashClaude Sonnet 5 · Grok 4.20GLM 5.2
EfficientVolume and latency.Gemini 3.5 Flash-LiteClaude Haiku 4.5 · Grok 4.1 FastGemma 4 26B · Qwen3-Next-80B Instruct
SpecialistOne job, done well.Gemini 3.1 Flash Image · Gemini Omni 1.1 Flash · Veo 3.1 · Lyria 3 · Gemini Embedding 2Codestral 2DeepSeek-OCR · MiniMax M2

Choose by

  • Intelligence
  • Latency
  • Modality
  • Cost
  • Sovereignty
  • Customisation
  • Deployment requirements

Deployment options differ by model: partner models are managed APIs; open models are managed or, where supported, self-deployed. Items marked Preview are pre-GA. Also available but not featured: Llama 4 Maverick and Scout (Apr 2025), Qwen3 235B and Qwen3 Coder (2025), gpt-oss 120B and 20B (Aug 2025), Mistral Medium 3, Small 3.1 and OCR (2025), AI21 Jamba 1.5 (2024), and earlier Claude, Gemini and Gemma generations.

Proof 02 · Floor 5

Don't just buy products.Build new ones.

Google Cloud meets developers where they are rather than forcing one development approach.

No / low code

Agent Studio

A low-code visual canvas for designing, prototyping and managing agent reasoning loops and workflows.

Code

Agent Development Kit

An open-source, model-agnostic framework for building and deploying complex agents. Use models and frameworks beyond Google's own.

An agent, assembled

Model
Instructions
Tools
Data
Memory
Other agents
= An agent that does a job

Associates working together

Multi-agent systems: specialised agents collaborate rather than one agent doing everything.

Agent Garden

Ready to use. Ready to customise.

A library of prebuilt agents and templates. Start with proven building blocks instead of rebuilding common capabilities from scratch.

  • Research pattern

    Gathers, reads and synthesises sources into a brief.

  • Data analysis pattern

    Answers questions against enterprise data and explains the result.

  • Customer conversation pattern

    Handles a service conversation and hands off when needed.

  • Workflow pattern

    Runs a multi-step task across tools with checkpoints.

Patterns are illustrative categories, not a product list.

Proof 03 · Floor 3

Models know the world.Your data teaches them your business.

The store knows your business because the store keeps your context.

BigQueryDatabasesDocumentsEnterprise searchGoogle WorkspaceMicrosoft 365SaaS systemsAPIsGROUNDINGAgent Platformsearch · RAG · connectorsYour agentscontextual answers

What the store remembers

  • Grounding
  • Enterprise search
  • RAG
  • Operational data
  • Analytics
  • Contextual responses

A model that knows the world is a good start. An agent that knows your customers, your policies, your inventory and your history is the product. Grounding is how the store keeps that context in the building and out of the generic answer.

Your enterprise data grounds your agents. It is not used to train public Google models.

Proof 04 · The column

Open at every layer.

Doors, corridors and standard interfaces. The point is optionality, not a slogan.

Openness is only worth something if it exists at every layer. A store with an open model floor and a locked front door is still a locked store. Doors, corridors and standard interfaces connect every department to every other.

The structural column

Identity · Security · Governance · Open Standards run through every floor. Openness and control are the same column, not opposite walls.

  1. ModelsGoogle, partner and open models on one floor
  2. AgentsAgents built here or sourced from partners; A2A between them
  3. FrameworksADK is open source and model-agnostic
  4. ToolsMCP tools, APIs, connectors
  5. DataBigQuery, databases, documents, Workspace, Microsoft 365, SaaS
  6. InfrastructureTPUs and GPUs; open software; flexible consumption
  7. EcosystemPartners, Marketplace, implementation services

Proof 05 · Floor 1

Autonomy without anarchy.

Switch to the X-ray. Every agent has an identity, every interaction has a controlled path, every interaction can be protected.

ROOFTOPPeopleFLOOR 5AgentsFLOOR 4ModelsFLOOR 3Data + ToolsFLOOR 2Build + RuntimeFLOOR 1Govern + OptimizeFOUNDATIONAI HypercomputerIDENTITY • SECURITY • GOVERNANCE • OPEN STANDARDS
Access badges

Agent Identity

Every agent has an identity.

Authentication and granular authorisation for agents, so the platform knows who is acting and what it may touch.

Security checkpoints and controlled doors

Agent Gateway

Every interaction has a controlled path.

A central policy enforcement point for user-to-agent, agent-to-tool and agent-to-agent interactions.

Loss prevention

Model Armor

Every interaction can be protected.

Guardrails against prompt injection, harmful content and sensitive-data leakage, applied to prompts and responses.

  • Authentication
  • Authorisation
  • Policy enforcement
  • Controlled access to tools
  • Prompt-injection protection
  • Sensitive-data protection
  • Enterprise guardrails

Proof 06 · Floor 1, continued

Security is the store’s job.Not the brand’s.

A department store does not ask each brand to bring its own guards. It has a maintenance crew that fixes the broken lock before anyone finds it, and a watch at the door, on the floor and in the control room. On Google Cloud that is CodeMender plus AI threat detection, and both apply whichever model you pick.

Why now

  • In May 2026 Google Threat Intelligence Group reported the first zero-day exploit it believes was written with AI, found in use against a widely deployed admin tool.
  • The same report describes state-backed groups using models for vulnerability discovery, and supply-chain compromises that reached AI gateway libraries and package registries.
  • Google’s reading: attackers now find and weaponise flaws at machine speed, so defenders need to find, prove and fix them at machine speed too.
The maintenance crew

CodeMender

Public Preview · July 2026 · limited customers

Finds the flaw, proves it is real, writes the fix. You approve it.

CodeMender is a code-security agent from Google DeepMind, now hosted on Gemini Enterprise Agent Platform. It wraps a security-tuned harness around Gemini and works on the code you build in the store: agents, tools and the applications around them.

  • Source stays in your environment: the CLI sends snippets, findings and patches, never the repository; nothing is used to train models.
  • C/C++, Go, Java, Python, TypeScript/JavaScript, Rust and Ruby, with common frameworks.
  • Runs on current Gemini Flash and Pro models; you choose the model per run for cost, speed or depth.
  • Also the remediation stage of Google AI Threat Defense (May 2026), alongside Wiz, Mandiant and Google Security Operations.
  1. Find

    Scans a repository for memory-corruption, injection, web, cryptographic and data-handling flaws, or imports findings from scanners you already run, including Wiz.

  2. Verify

    Builds the code and runs a proof-of-concept exploit in a sandbox you manage. Only exploitable findings go forward, which cuts false positives.

  3. Fix

    Generates a patch, tests it, and has a second model judge that behaviour is unchanged. The result is a diff in your CLI, IDE or CI pipeline for a developer to review and approve.

Announced by Google DeepMind on 6 October 2025 as a research agent; DeepMind reported 72 security fixes contributed to open-source projects in its first six months. Preview terms apply: supervise it, and keep a human on every change.

The watch

AI threat detection, three places at once

CONTROL ROOMAI Protection · Security OperationsTHE FLOORAgent Platform Threat DetectionTHE DOORModel Armor on Agent Gatewayone request, watched all the way up
  1. At the door

    Generally available

    Model Armor on Agent Gateway

    Every prompt, response and tool call that passes through the gateway is screened for prompt injection, jailbreaks, harmful content and sensitive-data leakage. Block, redact or log; violations surface in Security Command Center.

  2. On the floor

    Preview · announced April 2026

    Agent Platform Threat Detection · Agent Anomaly Detection

    A watcher sits beside each agent in Agent Runtime and flags malicious binaries, libraries or skills, reverse shells, container escapes and credential hunting; control-plane rules catch agent-initiated data exfiltration and suspicious token generation. Anomaly detection adds statistical models and a model-as-judge to flag reasoning that does not look like the agent’s normal behaviour.

  3. In the control room

    Generally available · March 2026

    AI Protection in Security Command Center

    An inventory of every model, agent, endpoint, data source and MCP server; CVEs and plaintext secrets in agent workloads; attack-path simulation with agents as high-value assets; over-privileged agents flagged. Findings appear in the Agent Platform’s own Security tab, and Google Security Operations agents triage and hunt across them.

What it means for the model choice

Change the brand. Keep the guards.

None of this is tied to a model. Swap Gemini for Claude, or Claude for an open model, and the door, the floor and the control room stay exactly where they were. Security is a property of the store, which is why adding a model never means adding a risk review.

Availability is as published by Google Cloud on 5 September 2026. CodeMender is a pre-GA offering for evaluation, not production, and its Gemini 3.5 Flash Cyber variant is limited to selected governments and partners. Figures are Google’s own, dated and attributed in Sources.

Proof 07 · Floors 2 and 1

AI you can actually operate.

Access to models is not an enterprise AI system. Enterprises need a production system around them.

From prototype to production

  1. 01Idea
  2. 02Build
  3. 03Test
  4. 04Deploy
  5. 05Scale
  6. 06Operate
Floor 2

Agent Runtime

Managed execution from prototype to production: fast provisioning, scaling, and persistent memory for long-running agents.

QUALITYCONTROLAgent executestaskTracecapturedPerformanceevaluatedFailureidentifiedAgentoptimisedRepeat

Don't just deploy AI.Know how it's performing.

Observability

What did the agent do?

Evaluation

Was it any good?

Quality control is the reason the model floor can carry rivals without fear. The store measures the product, not the brand.

Proof 08 · Foundation

The engine room.

Everything customers don't need to see, but depend on.

THE STORE ABOVETPUsGPUsNETWORKING · STORAGE · SERVING
Foundation

AI Hypercomputer

Google operates the infrastructure underneath the platform so enterprises can optimise around performance, economics and the character of each workload.

  • TPUs
  • GPUs
  • Networking
  • Storage
  • Serving infrastructure
  • Open software
  • Flexible consumption

So enterprises can optimise for

  • Performance
  • Latency
  • Scale
  • Economics

The best store still needs a great supply chain.

Stores within the store

Stores within the store

Third-party models, third-party agents, enterprise software and implementation partners participate through Google Cloud Marketplace and the partner ecosystem. Google Cloud can create an ecosystem without requiring everything to be built by Google.

  • Third-party models
  • Third-party agents
  • Enterprise software
  • Implementation partners
  • Marketplace

Proof 09 · Four archetypes

Which kind of AI storeare you building?

Four ways to construct an enterprise AI stack. None is wrong. Each optimises for something different.

The Boutique

Exceptional products. One primary brand.

Strength
Deep vertical integration.
Trade-off
Your platform strategy becomes more dependent on one provider.

The Bazaar

Maximum choice. Assembly required.

Strength
Breadth.
Trade-off
The customer owns integration, security, operations and governance.

The Warehouse

Everything you need to build it yourself.

Strength
Control and infrastructure flexibility.
Trade-off
Significant engineering required above the infrastructure.

The Department Store

CHOICE WITHOUT CHAOS.

  • Integration
  • Choice
  • Control
  • + Governance
Strength
Choice + integration + governance + common operations.
The offer
Choose the best product for every job. Let the platform make everything work together.
That is the position to own.

You don't know which modelwill win tomorrow.

You shouldn't have to.

Build your AI strategy around choice, not prediction.

Google Cloud

The department store for enterprise AI.

Choose the best capability for every job — without rebuilding the store every time the market changes.