Gemini — Chat & Setup
Gemini is Google’s AI assistant — and the front door to the whole Google AI track. This page covers what Gemini is, the model tiers and famously large context window you’ll be choosing between, the many surfaces Gemini shows up in (the app, AI Studio, the API, the CLI, and all across Workspace), and the handful of setup and privacy choices that make it far more useful than a plain chat box.
Gemini is a friendly robot from Google that you talk to by typing. It can read really long things — like a whole stack of books at once — and it lives inside the Google apps you already use, like Gmail and Docs. You can give it a big pile of files, build a little custom helper called a “Gem,” and ask it to go research something and write you a report. It’s one helper that pops up in lots of doorways.
What Gemini is — and the model family
☺ Like you’re 10: Gemini comes in a few sizes, like picking a delivery vehicle: a big powerful truck for the heaviest loads, a quick everyday van for most trips, and a tiny scooter when you just need something small delivered fast and cheap.
Gemini is Google DeepMind’s family of large language models. Two things set it apart from the pack: it was built natively multimodal from the start (text, images, audio, and video are first-class inputs, not bolted-on extras — see Multimodal), and it ships with an unusually large context window. Just like Claude’s Opus/Sonnet/Haiku, you pick a tier that trades speed and cost against raw capability:
| Tier | Best for |
|---|---|
| Pro | The most capable — hard reasoning, long agentic tasks, tricky code, and the deepest use of that big context. |
| Flash | The balanced workhorse — fast and strong for most day-to-day chat and building. |
| Flash-Lite | The quickest and cheapest — high-volume, latency-sensitive, cost-sensitive jobs. |
Recent generations also expose “thinking” behavior: the model can spend extra internal steps reasoning before it answers, and on the newer models you can dial that effort up or down (more thinking = better on hard problems, slower and pricier). That’s the same idea covered generally in Reasoning, just Google’s knobs. Version numbers move fast here — at time of writing the lineup is around the Gemini 3.x generation with Pro / Flash / Flash-Lite variants, and Pro’s context window runs to roughly a million tokens (with even larger windows appearing on the flagship). Treat any specific number as “true at time of writing” and check Google’s current model list. There’s also Gemma, Google’s family of smaller open models you can download and run yourself (that story lives in Local LLMs), and Nano, the on-device variant that ships inside Android and Chrome.
A million-token context is roughly thousands of pages — entire codebases, long videos, or a stack of PDFs — in a single prompt. It doesn’t replace retrieval for large, changing knowledge bases, but it changes what “just paste it in” can mean. More on that below.
Plans & access
☺ Like you’re 10: There’s a free ticket that already gets you in the door, a cheap monthly pass for more rides on the best rides, and a big pass for people who ride all day. And if your school or work already pays Google, you might have a pass without even knowing it.
You can start with Gemini for free — the free tier already includes daily access to the mainstream models, Deep Research, and Gems. Paid consumer plans (bundled under Google AI, billed through your Google account) raise usage limits, unlock the top Pro model and the biggest context, and add extras like cloud storage:
| Plan | What you get |
|---|---|
| Free | Everyday models, Deep Research, Gems, Canvas — with lower usage limits. |
| Google AI Pro | Higher limits, the top Pro model, the largest context in the app, and extra Google One storage. The typical “power user” pick. |
| Google AI Ultra | The highest limits and earliest access to new models and tools for the heaviest users. |
| Google Workspace | Gemini is bundled into Business/Enterprise editions, with admin and data controls — the version most companies use. |
Exact prices and tier names shift (a plan may be renamed or repriced between when this is written and when you read it), so confirm the current lineup. The important distinction to remember: consumer accounts (personal Gmail) and Workspace/enterprise accounts follow different data-handling rules — a point we return to in the privacy section. This mirrors the free → Pro → higher-tier ladder you saw for Claude and for GitHub Copilot; the difference is that Gemini rides on the same account you may already use for Gmail, Drive, and Android.
Many surfaces, one model
☺ Like you’re 10: The same helper shows up in lots of doorways — a chat window, a workshop for tinkering, a socket other programs plug into, and even quietly inside the apps you already use. One helper, many doors, and the door changes how much it does for you.
“Gemini” isn’t one app — it’s a model you reach through many surfaces, each a different point on the autonomy spectrum. Knowing which door fits the task is half the skill:
| Surface | What it’s for | Covered in |
|---|---|---|
| The Gemini app (this page) | Conversation, writing, analysis, quick code — with files, Gems, Canvas, and Deep Research. Web, mobile, and inside Chrome. | here |
| Google AI Studio | A free web playground to try prompts, tune parameters, compare models, and grab API code. The fastest way to experiment. | Gemini API |
| The Gemini API | Build your own app on Gemini — messages, tools, multimodal input, streaming, structured output. | Gemini API |
| Gemini CLI | An open-source agent in your terminal that plans and edits across your codebase. | Gemini CLI |
| Gemini Code Assist | AI completions and chat inside VS Code, JetBrains, and Cloud tooling — Google’s counterpart to Copilot in the editor. | Gemini CLI |
| Gemini in Workspace / Chrome / Android | The assistant embedded in Gmail, Docs, Sheets, Meet, the Chrome browser, and on your phone — help right where the work already lives. | here |
This is the autonomy spectrum again: the app keeps you in the loop turn by turn; the CLI and Code Assist hand off more; the API is where you decide exactly how much to delegate. Same shape as Claude’s (app → Claude Code → API) and Copilot’s (completions → chat → agents) — just Google’s surfaces. AI Studio is the odd one out: it’s not a product you ship, it’s the workbench where you figure out your prompt before you commit to code.
The features that actually matter
☺ Like you’re 10: These extras turn a chat box into a real helper: one can go read a hundred web pages and write you a report, one lets you build your own tiny specialist helper, one gives you a workbench to build things on, and its huge memory means you can hand it a giant pile of stuff all at once.
The free-text box is the least interesting part. These are the features that change how much Gemini can do for you:
- Deep Research. Give it a topic and Gemini browses many sources, reads them, and comes back with a structured, cited report — a multi-step research agent, not a single reply. Great for “survey these five vendors” or “summarize the state of X.”
- Gems. Custom, reusable assistants you configure once with instructions, tone, and (optionally) reference files — Google’s version of a saved specialist. The counterpart to Claude’s Projects/Styles and to a custom GPT.
- Canvas. A side workspace where Gemini’s substantial output — a document, code, a small app, a webpage — opens so you can iterate on it, run it, and export it. The analogue of Claude’s Artifacts.
- Native multimodality. Paste or upload images, PDFs, audio, and video and ask about them directly — the model reads them natively rather than through an add-on. (See Multimodal.)
- The very large context window. Because the window is so big, you can drop in whole documents, long transcripts, or entire small codebases and ask across all of them at once. This is Ellie-the-Elephant’s “hold it all in mind” approach — powerful when your material fits, and complementary to RAG when it doesn’t. It also ties directly to context & memory: a bigger window is more working memory per turn, but it’s not permanent memory, and it still costs tokens on every call.
- Connectors & Workspace grounding. Gemini can reach into Gmail, Drive, Calendar, and other Google apps (and external connectors) to ground answers in your data. Under the hood this is the same retrieval-and-tools plumbing you met in MCP and RAG — the tightest of these live right inside Workspace.
A million-token context is huge per conversation, but it isn’t long-term memory and it isn’t free — everything in the window is re-read (and re-billed) each turn. For large, changing, or private knowledge, retrieval is still usually the right tool; the giant window shines for one-shot analysis of material that fits.
Delphi the Dolphin: Hi Foxy — today I’m wearing my Gemini hat. Hand me whatever you’ve got.
Foxy: Whatever I’ve got? I’ve got the entire handbook — hundreds of pages, plus a two-hour video. Should I split it up?
Ellie the Elephant: No need! That huge context window means Gemini can hold the whole stack in mind at once — a big context is just a lot of working memory. Paste it all in and ask across it.
Professor Owl: And pick the right tier: Pro when you want it to reason hard over that whole pile, Flash for everyday speed, Flash-Lite when you’re running a lot of small jobs cheaply.
Timmy the Turtle: I checked the summary against the handbook — every claim traces back to a real page, and the window held it all. Approved.
Prompting Gemini well
☺ Like you’re 10: The clearer you ask, the better you get back. Say who Gemini should pretend to be, what you actually want, show one example, and drop the file first so it reads before it answers.
The Prompting lesson’s fundamentals carry over completely — be specific, give a role, show an example, ask for the format you want. A few habits pay off especially well with Gemini:
- Put the big stuff first, the question last. When you paste a long document or a batch of files, lead with the material and end with a clear instruction. It helps the model attend to what matters in a large context.
- Lean into multimodality. Don’t describe the chart — paste the chart. Don’t transcribe the meeting — drop the recording. Gemini reads them directly, so give it the real artifact.
- Save the pattern as a Gem. If you keep giving the same instructions, bake them into a Gem once instead of re-typing them — the reusable-instructions idea from Claude Projects, Google-flavored.
- Ask it to think when it’s hard, and to be terse when it’s easy. For thorny problems, invite step-by-step reasoning (or raise the thinking effort where the model exposes that control); for quick answers, ask for just the result.
- Ground it, then ask for citations. When answering from your files or the web, tell Gemini to use only the supplied sources and to point at them — the same anti-hallucination move from RAG.
Data controls & privacy
☺ Like you’re 10: There’s a setting that decides whether Gemini keeps a diary of your chats and whether a person might peek at some to help teach the robot. You can turn that off. And a chat you have for school or work follows stricter rules than one on your home account.
This is the part worth understanding before you paste anything sensitive, because consumer and enterprise accounts are governed differently:
- Consumer accounts (personal Gmail). Your Gemini Apps Activity setting controls whether conversations are saved and whether a sampled subset may be reviewed by humans and used to improve Google’s models. You can turn it off in your Google Account (Data & Privacy → Gemini Apps Activity), delete past activity, and note that reviewed chats can persist separately for a period — so don’t paste secrets you’d regret being sampled.
- Workspace & enterprise accounts. Under Google’s enterprise terms, your prompts and content are not used to train models outside your organization, and admins get controls over retention (auto-delete windows) and which Workspace services Gemini may search. This is the version to use for company data.
- API & Vertex AI. Paid API usage follows the developer/enterprise data terms rather than the consumer ones — important when you’re building. Details live in Gemini API.
Exact retention periods and toggle names change, so check the current settings — but the mental model is stable: know which account you’re in, and check the training/retention toggle before sharing anything you wouldn’t want sampled. The broader treatment of these trade-offs is in AI Security and Responsible AI.
The single most common mistake is pasting confidential material into a personal Gemini with activity on. For anything sensitive, use a Workspace/enterprise account (covered by no-training terms) or turn Apps Activity off first.
When to reach for Gemini
☺ Like you’re 10: Pick the helper that fits the job. Gemini is great when you have a huge pile to read at once, or when your stuff already lives in Google apps. Other helpers are great too — the trick is matching the tool to the task, not picking a favorite forever.
Gemini, ChatGPT, and Claude are all excellent general assistants, and they leapfrog each other constantly. Rather than crowning a winner, match the tool to the task. Gemini’s distinctive strengths:
| Reach for Gemini when… | Why |
|---|---|
| You need to feed in a lot at once | The very large context window handles whole codebases, long videos, or a stack of documents in one prompt. |
| Your work lives in the Google ecosystem | Tight, native integration with Gmail, Docs, Drive, Sheets, Meet, Chrome, and Android — help right where you already work. |
| The task is heavily multimodal | Native handling of images, audio, and video, not an add-on. |
| You want built-in Deep Research | A ready-made research agent that surveys many sources and writes a cited report. |
By contrast, you might prefer Claude (see Claude — Chat & Setup) when you value its careful, writing-forward style or its agentic coding surface, or ChatGPT for its own ecosystem and habits. The genuinely useful skill isn’t loyalty — it’s knowing each tool’s edges. For the underlying, provider-neutral concepts behind all of them, the AI landscape and How models work lessons are the shared foundation; the ecosystem and further reading pages map the wider field.
(1) Name Gemini’s three main model tiers and when you’d pick each. (2) What does a “million-token context window” let you do — and why is it not the same as permanent memory? (3) List three surfaces Gemini shows up in beyond the chat app. (4) Before pasting a confidential document, what one setting (and which account type) should you check first? Next up: Gemini CLI, where Gemini becomes an agent in your terminal, and Gemini API for building on it.
Check your answers
- The three tiers: Pro when you want it to reason hard over a large pile of material, Flash for everyday speed, and Flash-Lite when you're running lots of small jobs cheaply.
- The million-token window: It lets you drop in whole documents, long transcripts, or entire small codebases and ask across all of them at once. It's not permanent memory because it's working memory per conversation only — everything in the window is re-read (and re-billed) each turn, so it isn't retained long-term or free.
- Three surfaces beyond the app: Google AI Studio (a web playground for experimenting), the Gemini API (building your own app on Gemini), Gemini CLI (an agent in your terminal), Gemini Code Assist (completions and chat inside your editor), and Gemini in Workspace/Chrome/Android (embedded in Gmail, Docs, Meet, and more) — any three qualify.
- Check before pasting: Check the Gemini Apps Activity training/retention toggle on a consumer (personal Gmail) account, since that setting governs whether chats are saved and may be sampled by humans. For anything sensitive, prefer a Workspace/enterprise account (covered by no-training terms) or turn Apps Activity off first.