Before you start
- A Google account (personal, or Workspace for professional use)
- To understand how the models work and their limits, see Generative AI explained simply
In plain words
Gemini is both the name of Google's AI models and that of its assistant. Around them, Google has built a whole family of tools: the everyday assistant (the Gemini app), a research notebook over your documents (NotebookLM), a workshop for developers (AI Studio), an environment where agents write code for you (Antigravity), AI in Gmail and Docs (Workspace), and free models to install yourself (Gemma). This page gives you the full tour: what each tool is for, and which one to choose.
Information checked on 20 October 2026. Google renames and evolves its products very quickly: check model and menu names when you get started. No prices are given here.
1. The ecosystem in one table
| Tool | For whom | What it does |
|---|---|---|
| Gemini app | Everyone | Conversational assistant: questions, writing, in-depth research, images, actions in Google apps. Web, Android, iOS, macOS, Windows. |
| NotebookLM | Students, researchers, professionals | Answers only from your sources (PDFs, documents, websites, videos), with citations; generates audio overviews, study guides and mind maps |
| Gemini in Workspace | Businesses | AI built into Gmail, Docs, Sheets, Meet, Drive |
| Google AI Studio | Developers, the curious | Test models, tune parameters, create an API key, generate a complete application (Build mode) |
| Gemini API | Developers | Integrate the models into your programs |
| Antigravity | Developers | "Agent-first" development platform: desktop app, command line and development kit |
| Gemini Enterprise and Agent Platform (formerly Vertex AI) | Businesses | Enterprise agents, governance, model hosting in Google Cloud |
| Gemma | Homelab, developers | Open models, to download and run on your own machine |
| Flow, Omni, Nano Banana, Lyria | Creative work | AI-generated video, images and music |
2. The models
| Family | Models (October 2026) | Use |
|---|---|---|
| Gemini 4 | Gemini 4 Argon: announced on 30 September 2026 for cybersecurity partners, opened on 13 October in the API and to AI Pro and Ultra subscribers | The most capable: code, long agent tasks, maths, science |
| Gemini 3.x Flash | Gemini 3.8 Flash (September 2026), 3.6 Flash, 3.5 Flash-Lite | Fast and economical; Flash is the everyday default model, Flash-Lite for large volumes |
| Specialised | 3.8 Live (real-time voice conversation), 3.5 Transcribe (transcription with speaker diarisation), 3.8 Flash TTS (speech synthesis) | Voice, meetings, subtitles |
| Image | Nano Banana 2.1 | Image creation and editing, up to 4K |
| Video | Gemini Omni, which succeeds the Veo models in the API; available in Flow | Generating and extending videos |
| Music | Lyria 3.5 | Complete stereo tracks |
| Open | Gemma 4 (April 2026) | See section 9 |
What characterises the Gemini models: a very large context window (whole documents, hours of video), native processing of images, audio and video, and access to Google Search to ground answers in recent information.
3. The Gemini app
Where to use it
- Web: gemini.google.com
- Android: Gemini replaces the old Google Assistant; long-press the power button or say "Hey Google"
- iOS, macOS (native app), and Windows since September 2026:
Alt + Spaceopens Gemini on top of any window - In Chrome, Gmail, Docs and the other Google apps, through the Gemini button
The main features
| Feature | What it does |
|---|---|
| Model choice | A fast model by default, and Gemini 4 Argon for complex tasks (Pro and Ultra plans) |
| Deep Research | In-depth research: Gemini consults dozens of sources and returns a structured, sourced report, exportable to Docs |
| Skills | Saved, reusable instructions that can be combined; they replace Gems, which are migrated automatically |
| Customize | Left-hand menu: how it answers, skills, and connecting more than 40 apps (Gmail, Drive, Calendar, and third-party services) |
| Personal Intelligence | Personalisation option: Gemini draws on your connected Google apps for answers suited to your context |
| Gemini Live | Natural voice conversation, with camera or screen sharing on mobile |
| Images | Creating and editing with Nano Banana, directly in the conversation |
| Notebooks | Organising conversations by theme, synced with NotebookLM |
| Temporary chat | A conversation not kept in the history |
| Antigravity on macOS | Hand over a coding task from the conversation, and be notified when the code is ready to review |
| Gemini Spark | An agent that works continuously across your Google apps (emails, documents, spreadsheets); for Ultra subscribers in the United States only |
The plans
| Plan | What it brings |
|---|---|
| Free | The core of the assistant with the fast model, lower usage limits |
| Google AI Plus | Higher limits, Gemini Omni and Gemini 3.5 Flash |
| Google AI Pro | Everything in Plus, access to Gemini 4 Argon, extra credits for Antigravity and Flow, Google perks (YouTube Premium Lite, etc.) |
| Google AI Ultra (two tiers) | 5 to 20 times Pro's limits in Gemini and Antigravity, priority access to Antigravity, 20 TB of storage, Gemini Spark (United States) |
Since May 2026, limits are no longer a number of messages per day: they depend on the computing power consumed (model size, conversation length, features used), refill roughly every five hours, with a weekly cap. Beyond that, Gemini switches to a lighter model.
4. NotebookLM: the AI that only talks about your documents
NotebookLM works differently from the assistant: you create a notebook, add sources to it (PDFs, Google Docs, web pages, YouTube videos, audio files), and the AI answers only from those sources, with clickable references to the passage cited.
| Use | Example |
|---|---|
| Querying documentation | Load a firewall's administration guide and ask "how do I turn on IPS?", with the exact passage cited |
| Preparing a meeting | Load the minutes and the specification, ask for a status update |
| Learning | Generate an audio overview (two voices discussing the subject), a video, a mind map, flashcards or a quiz |
| Training a team | Share the notebook with your colleagues |
It's Google's most reliable tool for working on documents, because it greatly limits invention: it tells you when the information isn't in the sources. For a team knowledge base, see also knowledge bases.
5. Google AI Studio: the developers' workshop
Google AI Studio is the web interface for testing models and building with them.
| Feature | Use |
|---|---|
| Playground | Try a model, compare two models side by side, set system instructions, thinking level, structured output (JSON) |
| Tools | Turn on grounding with Google Search, code execution, URL reading |
| Get API key | Create a key for the Gemini API, attached to a Google Cloud project |
| Get code | Export the tested request in Python, JavaScript or curl |
| Build | Describe an application in natural language: an agent (Antigravity's) generates a web app (React on the client, Node.js on the server) or an Android app (Kotlin, Jetpack Compose) |
In Build mode, the API key is stored as a server-side secret (GEMINI_API_KEY variable, Secrets panel) and never appears in the code sent to the browser. The application deploys in one click to Cloud Run (Google Cloud billing by usage), syncs with GitHub, or downloads as a ZIP. Personal applications such as a family planner or an internal tool can be prototyped this way in a few minutes, provided you review the code before hosting them.
A first API call from a terminal:
export GEMINI_API_KEY="your-key"
API="https://generativelanguage.googleapis.com/v1beta/models"
curl "$API/gemini-3.8-flash:generateContent" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-H "Content-Type: application/json" \
-d '{"contents":[{"parts":[{"text":"Explain NAT in three sentences."}]}]}'
Free, but at what cost to your data?
On the free tier of the API and AI Studio, Google may use your prompts and responses to improve its products, and people may review them (after disconnecting them from your account). Google asks you not to send sensitive or personal information. On the paid tier, data isn't used for training. Notable exception: for users in the European Economic Area, Switzerland and the United Kingdom, the paid tier's terms apply, even to free use. Check the current terms before processing business data there.
Never publish an API key in a Git repository or in a web page's code: keep it in a secrets manager or an environment variable, and restrict it in the Google Cloud console.
6. Antigravity: coding with agents
Google Antigravity, launched in November 2025 and moved to version 2.0 in May 2026, is Google's development platform designed for agents, rather than for an assisted human.
| Element | Role |
|---|---|
| Desktop app (Windows, macOS, Linux) | Based on VS Code. Editor view (classic editor with an agent in the side panel) and Manager view (several agents working in parallel on several projects) |
| Artifacts | Each agent produces verifiable items: task list, implementation plan, screenshots, browser recordings. You review the plan before the code. |
| Built-in browser | The agent opens the application it's developing, clicks, checks the result |
| Antigravity CLI | The command-line version, which replaced Gemini CLI for individuals on 18 June 2026. Skills, hooks, subagents and extensions (now "plugins") are kept. |
| SDK | For building and hosting your own agents |
| Models | Gemini (Flash and the most advanced models), and also third-party models, including Anthropic's Claude and the open GPT-OSS model |
Businesses with a Gemini Code Assist Standard or Enterprise licence keep Gemini CLI and the extensions for VS Code and JetBrains.
A coding agent has access to your machine
An agent that runs commands and browses can also delete files or expose secrets. Work in a clean Git repository so you can undo everything, keep manual approval of terminal commands, don't leave production .env files in the project, and prefer a virtual machine or a container for experiments. The same principles apply to Claude Code.
7. Gemini at work: Workspace and the enterprise
- Gemini in Workspace: summarise a Gmail thread, draft a reply, generate a table in Sheets, take meeting notes in Meet, query Drive content. For work and school accounts, Google applies terms distinct from those of personal accounts: point your users to the Workspace generative AI privacy hub and check your contracts.
- Gemini Enterprise: the agent platform for the whole organisation, connected to internal data (Drive, SharePoint, Confluence, CRM…), with governance and permissions.
- Gemini Enterprise Agent Platform (formerly Vertex AI): the models in Google Cloud, with choice of region, logs and enterprise contracts. It's the route to favour for a business application processing sensitive data. See also the cloud comparison.
Workspace administrators control whether Gemini is turned on per organisational unit in the admin console. As with any cloud service, ask about data location and reversibility (see On-premises or SaaS).
8. Privacy: the settings to check
In the Gemini app, Settings › Activity menu:
| Setting | Effect |
|---|---|
| Keep Activity | On: your conversations are saved and may be used to improve Google services, including training, with human review of some of them. Off: conversations are kept for 72 hours then deleted, and aren't used for training. |
| Auto-delete | 18 months by default; can be set to 3 or 36 months |
| Reviewed conversations | Some conversations are reviewed by people, disconnected from the account, and kept for up to 3 years, even if you delete your activity |
| Live audio and video | A separate setting, off by default for service improvement |
| Connected apps | Gmail, Drive, Calendar… connect them only if needed. Some features (Workspace apps, tasks handed to Gemini) require activity to be kept. |
The simple rule
Never paste into a consumer assistant what you wouldn't put in an email to a stranger: passwords, customer data, complete configurations with public addresses. For sensitive data, use a business plan with contractual guarantees, or a local model (section 9).
9. Gemma: the open models
Gemma 4, released in April 2026 under the Apache 2.0 licence, can be downloaded and runs on your own hardware, without sending any data to Google.
| Model | Size | Inputs | Indicative hardware |
|---|---|---|---|
| Gemma 4 E2B | 2.3 billion effective parameters | Text, image, audio | Raspberry Pi 5, phone |
| Gemma 4 E4B | 4.5 billion effective | Text, image, audio | 12 to 16 GB of memory, Apple silicon Mac |
| Gemma 4 26B (MoE) | 3.8 billion active | Text, image, audio, short video | About 24 GB |
| Gemma 4 31B | 30.7 billion | Text, image, audio, short video | 40 GB or more |
They install with Ollama and are used with an interface such as Open WebUI: see Ollama and Open WebUI in a homelab.
ollama pull gemma4:e4b
ollama run gemma4:e4b "Summarise the differences between RAID 1 and RAID 5."
The exact model name in Ollama may vary: check it in the Ollama library.
10. Gemini, Claude or another?
| Need | Gemini's strength | Other assistants' strength |
|---|---|---|
| Living in the Google ecosystem (Gmail, Drive, Android) | Native integration everywhere | — |
| Working on your documents with citations | NotebookLM | Claude's projects |
| Images, video, music | Nano Banana, Omni, Lyria | Fewer built-in tools elsewhere |
| Video and audio as input | Very good native support | Varies by assistant |
| Coding with agents | Antigravity | Claude Code, and the tools built into IDEs |
| Long writing, reasoning, complete tasks | Gemini 4 Argon | Claude, known for writing and working autonomously |
| European sovereignty | — | Mistral AI, a French publisher |
Many users combine two tools: the one built into their working environment, and a second for the tasks where it excels.
Key points
- Gemini means both the models and the assistant; the ecosystem also includes NotebookLM, AI Studio, Antigravity, Workspace and Gemma.
- The models change every month: Gemini 4 Argon has been the most advanced since October 2026, the Flash models are for everyday use.
- NotebookLM is the best choice for working on your own documents.
- AI Studio is for testing models, creating an API key and prototyping an application; watch out for the free tier's data use terms.
- Antigravity replaces Gemini CLI for individuals: agents that plan, code and test, to be watched closely.
- Check the Keep Activity setting, and keep sensitive data for business plans or local Gemma models.
Sources
- Google: Gemini 3.5, presentation (May 2026)
- Google: Gemini API release notes
- Google: Gemini app release notes
- Google: Google AI subscriptions after I/O 2026
- Google: AI Studio, Build mode
- Google: Gemini API terms (data use)
- Google: Gemini Apps privacy
- Google Developers: from Gemini CLI to Antigravity CLI
- Google: official Antigravity website
- Google DeepMind: the Gemma models (Gemma 4, April 2026)
- Google: Gemma 4 documentation and its licence
- Google: Gemini 4 Argon, announcement and roll-out to cyber defence specialists (30 September 2026)
- Google France: Gemini 4 Argon available (13 October 2026) (in French)