Gemini Complete Guide: Plans, Models, Skills, AI Studio, and Antigravity

About This Guide

This guide is a reference document for Gemini, originally written in Traditional Chinese. It pulls together information scattered across official Google pages into one place: which products exist, how the plans are divided, how to choose a model, how to connect to the API, how to use the command-line development tool, and how your data is handled.

Query date and how to use this guide

All facts were queried on 2026-11-06. Each chapter ends with its sources and query date. Anything in the guide without a citation follows the official pages as of that day.

Gemini changed a great deal this past year. Names, plans, models, and prices have almost all been replaced. The prices, dates, and feature scope in this guide may change again. Check the official pages once more before you make a decision.

What changed this year

If your picture of Gemini is still stuck in 2025, update these points first:

Writing principles

• This guide only states facts you can verify on official pages. Where something cannot be found, or official pages contradict each other, it says “not published by Google” or “pages disagree” directly, and it does not fill the gap with third-party numbers.
• Prices in NT$ are labeled as reference prices. Go by what the official Google page shows at the time.
• It does not rank models by quality. Each vendor’s benchmarks are published by that vendor and cannot be compared directly. Test with your own tasks.
• Some features in this guide have restrictions on age, region, or account type. Each chapter notes them where they apply.

How to read it

To decide which entry point to use, read Chapter 2. To decide whether to pay, read Chapter 3. To write code or connect to the API, read Chapters 10 and 11. If you are worried about where your data goes, go straight to Chapter 12.

Sources: Gemini release notes, Gemini API models page, Gemini subscription plans (queried 2026-11-06)

What Gemini Is: The Product Map

Inside Google, “Gemini” refers to both a model family and a whole line of products. Sort out the entry points first, then decide which one to use.

Entry points for general users

Google’s consumer plan page lists Gemini-related products as one product line:

• Gemini app: chat, file analysis, Canvas, Deep Research, Gemini Live, and image and video generation.
• Google Flow: image and video creation.
• Gemini Notebook: notebooks linked to NotebookLM (Chapter 5).
• Gemini in Gmail, Docs, and Vids: use it directly inside Google’s apps.
• AI Mode and agentic features in Search.
• Google Antigravity: for development.
• Project Genie: Ultra plan.
• YouTube Premium Lite and Premium: bundled with the Pro and Ultra plans.

On the desktop: the Gemini app for Windows launched on 2026-09-10, and the macOS version added the Antigravity coding interface on 2026-10-13.

Three entry points for developers

The Antigravity docs list four surfaces:

• Desktop app (v2.5 or later)
• CLI agy (v1.2 or later)
• IDE extensions (VS Code, JetBrains, Visual Studio, Zed, Xcode, Eclipse)
• Antigravity SDK (Python)

How to pick an entry point

Recent official updates timeline

Key dates from the official release notes over the past three months:

On the day of the query (2026-11-06), the release notes had no November entries yet.

Sources: Gemini subscription plans (Taiwan), Google Antigravity getting started, Gemini release notes (queried 2026-11-06)

Plans and Pricing

Gemini’s personal plans are now three tiers, Google AI Plus, Pro, and Ultra, plus a free tier. A plan decides more than the model. It also bundles storage, Flow credits, and other Google services.

Personal plans (Taiwan)

The table below shows the list prices on the Taiwan official page. NT$ amounts are reference prices. Go by the checkout page for the actual price:

The official page does not say clearly which Flow credits figure belongs to which of the two Ultra tiers.

Pro also includes: a 1M token window, Deep Research, a limited trial of Veo 3.1 Lite, YouTube Premium Lite, access to Antigravity’s basic agent models, and 5x the quota for things like Notebook audio overviews. Ultra also includes: YouTube Premium individual plan, Project Genie, 5x or 20x the Antigravity quota, and early access to Tasks in Gemini (limited to certain countries).

Official US prices

You can cancel a plan at any time. Some billing periods are not refunded (except where the law requires it). For student plan conditions, see the 2026-08-19 release note.

Can you buy it in Taiwan

In the country list of the official “Gemini Apps limits & feature availability” page, Taiwan appears on the availability lists for both AI Plus and for AI Pro and Ultra.

How usage is counted

• Usage is compute, not a fixed number of prompts.
• It resets every 5 hours, until you reach the weekly limit.
• Video generation uses the most, followed by Deep Research, and regular chat uses the least.
• You can buy AI credits for more capacity. Check your current usage under Usage Limits in Settings.
• After you use up your quota, you can still use Flash-Lite.
• Google says limits may change with capacity, without notice.

Context window (Gemini app)

Business and team plans

Google Workspace’s plan page also lists plans that include Gemini (AI chat, Notebook, and AI features in Gmail, Drive, Sheets, and Meet). Enterprise is priced on inquiry. Google has not published NT$ prices.

Antigravity plans

Every plan includes Gemini 3.8 Flash and unlimited Tab completion. Pro has a quota that resets every 5 hours plus a weekly limit, and Ultra is the highest. Pro and Ultra can buy AI credits to go over the quota (you can set it to Never or Always). Bring your own API key (BYOK) is not supported. Google has not published the exact quota numbers and only says “higher.”

Sources: Gemini subscription plans (Taiwan), Gemini subscription plans (US), Gemini Apps limits & feature availability, Antigravity Plans, Google Workspace AI (queried 2026-11-06)

The Model Family

Gemini entered its 4th generation in 2026. The models available on the API and their status are below. The table comes from the official models page (2026-11-06).

Models available on the Gemini API

Models for specific purposes: Nano Banana 2.1 (gemini-nano-banana-2.1), Nano Banana 2 (gemini-3.1-flash-image), Nano Banana 2 Lite, Nano Banana Pro (gemini-3-pro-image), Gemini Omni Flash, Gemini 3.8 Live, 3.5 Transcribe, various TTS models, and Gemini Embedding 2 (gemini-embedding-2).

Gemini 4 Argon

Google positions it as the flagship reasoning model, suited to complex software engineering, math and science, autonomous multi-step agents, and long decode.

Release timeline:

• 2026-09-30: Provided first to the Fairwind Program, which includes cybersecurity defenders.
• 2026-10-13: Opened in the Gemini app to AI Pro and Ultra (Google’s official blog).
• API: Available on paid tiers, not available on the free tier.

Pricing (USD, per million tokens):

In other words, the current price is a phased price and changes in December.

Other main models

Gemini 3.8 Flash: Google positions it as the strongest Flash, for long-running software engineering, agents, and enterprise workflows. Input 1,048,576 and output 65,536; thinking is low, medium, or high, and minimal is not supported and setting it returns an error. Price $0.75 / $3.75, and $1.50 / $7.50 from 2027-01-01. Available on the free tier.

Gemini 3.5 Flash-Lite: Low latency and low cost, for sub-agents and document parsing. Price $0.30 / $2.50, available on the free tier (no cache).

Gemini 3.1 Pro Preview: Input is $2 up to 200k and $4 above; output is $12 up to 200k and $18 above. Not available on the free tier.

Gemini 4 Carbon

The DeepMind page lists Gemini 4 Carbon as the next frontier model. At the time of writing it is not open to general users, and it does not appear on the API pricing page either. Google has not published the details.

Thinking settings

Thinking tokens and output tokens are billed together. The API only returns a summary of the thinking (thinking_summaries: "auto"), but you are billed for the full thinking. Google recommends running your own evaluation at each level and picking the lowest level that meets your bar.

Parameter changes

From Gemini 3.6 onward and with Gemini 4: remove temperature, top_p, and top_k from your requests. These parameters are currently ignored and will return an error in the future.

Models you can choose in Antigravity

Claude Sonnet 4.6, Opus 4.6, and GPT-OSS-120b are scheduled to be removed on 2026-11-18. Reasoning effort has three levels: Low, Medium, and High.

What you can choose in the Gemini app

The official support page says outright that it does not list models plan by plan. The subscription page words it as: the free tier has Flash-Lite, Plus has Flash, and Pro includes Argon, Pro, and Flash. The pages differ slightly, so go by the model menu in your own account.

Sources: Gemini API Models, Gemini 4 Argon, Gemini 3.8 Flash, Gemini 3.5 Flash-Lite, Pricing, Thinking, Gemini 3 developer guide, DeepMind Gemini models page, Gemini 4 availability, Antigravity Models (queried 2026-11-06)

Everyday Chat and Files: Canvas, Notebook, Deep Research

This chapter covers the workflows you will use most in the Gemini app, and the limits of each.

Canvas

Canvas is a workspace where you edit a deliverable together with Gemini. You can use it for:

• Writing and documents
• Code
• Apps and games
• Web pages
• Infographics
• Quizzes
• Audio overviews

All Gemini users can use it, and Pro and Ultra get a 1M window. Projects can be shared by link. On mobile you can see projects, but editing text formatting is only available on the desktop web version.

Deep Research

Deep Research first produces an editable research plan, and only starts searching after you confirm it.

Gemini Notebook

Since 2026-04-08, Gemini has “Notebooks” that sync both ways with NotebookLM: add a source on either side and it shows up on the other. You can add custom instructions and files. At the time of the announcement, users under 18 and Workspace and Education accounts were not supported. On the Pro plan, quotas such as Notebook audio overviews are 5x.

Google has not published numbers for the source limit or the free quota of Notebook.

Connecting Google Workspace

After you connect Google Workspace apps in Gemini, you can:

• Search and summarize Gmail, Docs, and Drive
• Manage Tasks, Keep, and Calendar

To use it, type @ and pick the service. Prerequisites:

• Keep Activity must be on (see Chapter 12).
• Work or school accounts need administrator permission.
• Gmail smart features must be on.

What it cannot do right now: read comments and images inside Docs and Gmail; manage Drive (create folders, move files); calculate or report storage space.

Small reminders

• For model choice and usage limits, see Chapters 3 and 4.
• Guided Learning (step-by-step guided study) launched in 2025-08 but can no longer be found on the official overview pages in 2026, so this guide does not list it as a core feature.

Sources: Canvas, Deep Research help, Deep Research overview, Connect Google Workspace apps, Notebooks announcement (queried 2026-11-06)

Skills and Gems

Gemini is replacing Gems with Skills as the way to “save instructions once and reuse them.” We are in the transition period now and both exist side by side, so this chapter focuses on Skills and only covers the Gems shutdown timeline and migration.

What Skills are

Google announced Skills on its official blog on 2026-09-30:

• Save an instruction you use often once, then type / plus its name to run it.
• You can create one directly from an existing conversation.
• You can stack multiple skills.
• Reference files are supported: text, PDF, and images.
• Available on every Google AI subscription tier.
• Currently limited to users 18 or older.
• Gemini Spark already supports Skills.

Skills are about to replace Gems. Google says it will later add what Gems already had: sharing, and support for Drive and Notebook files.

Gems shutdown timeline

• When Gems are retired, they are automatically converted to skills.
• Opal also shuts down in November.
• Gems in Google Labs will shut down, and will not be migrated.
• On the day of the query, official pages still listed Gems among the free plan features, which means the two exist side by side during the transition.

What to do

1. Keep your own backup of your existing Gems content (the instructions and the reference files).
2. Try / in a conversation and check that each skill’s name and content match what you expect.
3. For a Gem you need to share with others, first check whether the new Skills have gained sharing, then decide whether to rebuild it.

Skills in the command-line development tool

Antigravity has its own skills mechanism. It follows the agentskills.io open standard, and its format is different from the app version of Skills above. See Chapter 11 for that.

Sources: Skills official blog, Gems overview (queried 2026-11-06)

Personalization and Memory

Gemini has three mechanisms that help it “know you better”: connecting your Google data (Personal Intelligence), memory across conversations, and importing from other AI tools.

Personal Intelligence

Launched in beta on 2026-01-14.

The current official overview page describes a global beta that excludes the European Economic Area, Switzerland, the UK, and Nigeria. The page does not list Taiwan one by one, and Taiwan is not on the exclusion list, but the page requires you to “meet eligibility requirements,” so whether you can actually turn it on depends on your account settings.

Memory and import

The import feature launched on 2026-03-26:

1. Go to Settings and choose import.
2. Copy the prompt Gemini provides and paste it into your old AI app.
3. Paste the summary the old AI returns back into Gemini.
4. You can also upload a ZIP file to import your chat history directly.

Restrictions: available to all consumer accounts; Business, Enterprise, and users under 18 are not supported; not yet available in the European Economic Area, the UK, or Switzerland. The original “past chats” was renamed “memory.”

Gemini Live can also remember details across conversations, such as dietary preferences and important dates.

Privacy connections

Memory, Workspace connections, and the Keep Activity setting are related. See Chapter 12 for details. If you plan to turn off Keep Activity, know first that Gemini can no longer connect to Google Workspace once it is off.

Sources: Personal Intelligence announcement, Personal Intelligence overview, Memory import announcement, Gemini Live (queried 2026-11-06)

Tasks and Gemini in Chrome

This chapter covers Gemini’s agentic features: letting it get things done for you while you are not operating the computer. This is currently the area with the most changes and the least consistency between official pages, so go by what your own account shows before you use it.

Tasks mode

Google’s official announcement on 2026-10-13:

• The main chat window has a toggle that switches from chat to Tasks.
• Gemini can complete tasks across Google apps while you are away from your computer.
• Gemini connects to “40 or more apps,” such as Instacart, OpenTable, Adobe, and Shopify.
• Pro and Ultra subscribers can “assign tasks.”

Plans and regions: the pages disagree

Google has not given a consistent explanation of whether it is available in Taiwan yet or whether Pro includes it.

Spark

Spark is the background agent announced at Google I/O 2026 (2026-05-19). From the end of May it opened to US Ultra users aged 18 and older. The Skills announcement on 2026-09-30 still says “Gemini Spark already supports Skills,” but most official pages now only mention Tasks. Google has not said outright how the names Spark and Tasks relate to each other.

Gemini in Chrome

• Summarize tabs, compare multiple tabs, and use Gemini Live.
• Auto browse: currently a US-only preview, limited to Pro and Ultra.
• On Android it can work with what is on your screen; on iOS it is built into Chrome.
• Taiwan: Taiwan is in the official region menu, but the page does not clearly say it is available in Taiwan; Taiwanese media reported that it became available in Taiwan from 2026-04-21.

Connected apps

Connected apps added at the Made by Google event on 2026-08-12 (rolling out gradually): Granola, Otter.ai, Wix, Fever, GetYourGuide, Localiza, OpenTable (UK), Ticketmaster, iHeartRadio, Pandora, Angi, Thumbtack, Zocdoc.

Usage tips

• Agentic features operate your account or place orders for you. The first time you use them, start with low-risk tasks (looking things up, organizing an itinerary).
• Before you connect a new app, read what data it can access.

Sources: Gemini 4 availability (including Tasks mode), Gemini subscription plans, Gemini Apps limits & feature availability, Gemini in Chrome, New connected apps, Skills announcement, Taiwanese media report (technews) (queried 2026-11-06)

Voice, Images, and Video

Gemini Live

Tap the Live icon to talk with Gemini by voice:

• You can interrupt at any time and change topics.
• Voice and text live in the same conversation.
• It can show maps, weather cards, and photos.
• It can connect to Gmail, Drive, Calendar, Tasks, Keep, Maps, YouTube, and smart home.
• You can show it things with the camera.
• It can remember details across conversations.

Restrictions: Live cannot use Gems or Notebook, and does not support Omni or Lyria. The official pages do not list supported languages or screen sharing, and they do not say whether Traditional Chinese is supported.

Image generation

Every generated image carries an invisible SynthID and a visible watermark. You can upload a file and ask Gemini to verify whether it was generated by Google AI. You must be 18 or older to use it.

Video generation

• Video generation in the app now uses Gemini Omni 1.1 Flash, replacing Veo 3.1.
• 10-second videos with built-in voice.
• Supports image-to-video (up to 5 reference images), video editing, scene extension, and AI avatars.
• Requires a subscription.
• Google does not state the resolution or the daily limit.
• Videos also carry SynthID.

Video is the feature that uses the most quota (Chapter 3), so estimate how often you will use it before you subscribe.

API side

For the model IDs of the image and video models, see Chapter 4. The API prices of these models are outside the scope of this guide. See the official pricing page.

Sources: Gemini Live (Taiwan), Image generation, Video generation (queried 2026-11-06)

Google AI Studio and the Gemini API

Get an API key

Create one at aistudio.google.com/apikey. The environment variable the SDK reads is GEMINI_API_KEY or GOOGLE_API_KEY, and when both exist, GOOGLE_API_KEY takes priority.

Since 2026-05-28, newly created keys are all auth keys (bound to a service account). Unrestricted standard keys are rejected, while standard keys with explicit restrictions still work.

SDK and your first request

pip install -U google-genai
npm install @google/genai

The main interface has changed to the Interactions API:

from google import genai

client = genai.Client()
interaction = client.interactions.create(
    model="gemini-3.8-flash",
    input="Explain how AI works in a few words",
)
print(interaction.output_text)

• Server-side state is chained with previous_interaction_id.
• REST requests need the Api-Revision: 2026-05-20 header.
• Google also provides a gemini-api CLI.
• The official quickstart says the Interactions API replaces the old generate-content mode. Google’s pages do not say clearly when the old interface will be retired, so for existing code, confirm in the official docs first.

Free tier and paid tiers

The official pricing page marks these text models as free to use: 3.8 Flash, 3.6 Flash, 3.5 Flash-Lite, 3.1 Flash-Lite, 3 Flash Preview, and the 2.5 series. Gemini 4 Argon and 3.1 Pro Preview are not available.

The public pages do not give specific RPM and RPD numbers for the free tier. You need to sign in and look at the rate-limit page in AI Studio.

• Limits are counted per project, not per key.
• RPD resets at midnight Pacific Time.
• Search grounding defaults to 1,500 RPD per project.

Prepay and billing

Pricing (standard tier, USD per million tokens, input / output)

Batch prices are about half the standard price (for Argon, the current Batch price is $1 / $5). Prices for other models such as audio, image, TTS, and embedding are not included in this guide. See the official pricing page.

Argon in practice on the API

• Remove temperature, top_p, and top_k.
• max_output_tokens includes thinking. If an older app sets it at 32K or below, raise it to at least 32,768.
• Set the HTTP timeout to at least 15 minutes for hard problems.
• Long outputs need background: true, and explicit cache is not supported.

Other API tools

Search grounding, Google Maps grounding, code execution, computer use, file search, URL context, function calling, structured outputs, Batch, and Flex and Priority inference.

Using long context

• Put the question after the context.
• Use cache for content you reuse.
• Accuracy drops when you retrieve multiple targets at the same time.
• 1M tokens is roughly 50,000 lines of code, or 8 novels.

AI Studio Build mode

The official docs pages do not describe Build mode in detail, so this guide does not cover it.

Sources: API key, Quickstart, Pricing, Rate limits, Billing, Gemini 3 developer guide, Long context (queried 2026-11-06)

Antigravity and Antigravity CLI

Google’s agent development tools were consolidated into Google Antigravity in 2026. Gemini CLI has moved into Antigravity CLI (command agy), and Jules has been announced as ending. This chapter focuses on agy, and covers Gemini CLI and Jules at the end.

The four Antigravity surfaces

Desktop app (v2.5 or later), CLI agy (v1.2 or later), IDE extensions (VS Code, JetBrains, Visual Studio, Zed, Xcode, Eclipse), and the Antigravity SDK (Python). For plans and quotas see Chapter 3, and for the models you can choose see Chapter 4.

Install agy

# macOS / Linux
curl -fsSL https://antigravity.google/cli/install.sh | bash

# Windows PowerShell
irm https://antigravity.google/cli/install.ps1 | iex

# Verify
agy --version

The default install location is ~/.local/bin/agy.

There are three ways to authenticate:

In headless environments such as SSH, it prints an authorization URL and a 6-digit OTP. With an API key, the official settings page words it as modelProvider: "gemini" together with GEMINI_API_KEY.

Common commands

The slash commands available in a conversation include /model, /usage, /fast, /resume, /permissions, /mcp, /skills, /agents, /statusline, /goal, /grill-me, /schedule, /browser, and /plugin. The actual list changes with the version, so go by what you see after typing /.

Shift+Tab switches between auto, plan, default, and accept-edits modes; Ctrl+R views artifacts.

Settings file

The location is ~/.gemini/antigravity-cli/settings.json (permission 0600).

• Permission rules are split into allow, ask, and deny, with the priority deny > ask > allow, in a format like command(git status).
• Other common keys: model, modelProvider, agentMode, toolPermission, enableTerminalSandbox, allowNonWorkspaceAccess.
• Hooks, MCP, and plugins go in ~/.gemini/config/ or <workspace>/.agents/.
• The telemetry option is stored on the account, not in the settings file, and is always off for business accounts.

Rules files: GEMINI.md and AGENTS.md

Antigravity looks for AGENTS.md or GEMINI.md in every directory level. It also reads .agents/AGENTS.md, .agents/GEMINI.md, and .agents/rules/*.md.

Notes:

• Rules are cumulative, not overriding.
• Every .md file in rules/ needs frontmatter with a trigger, whose allowed values are always_on, model_decision, glob, and manual. A file in the wrong format is silently dropped.
• A single file over 24,000 bytes is truncated.
• Global rules and always_on rules share a 20,000-token budget.

MCP

The global config is ~/.gemini/config/mcp_config.json, and the workspace config is <workspace>/.agents/mcp_config.json:

{
  "mcpServers": {
    "sqlite-explorer": {
      "command": "node",
      "args": ["/path/server.js"],
      "env": { "K": "V" }
    },
    "my-remote-server": {
      "serverUrl": "https://api.example.com/mcp/",
      "headers": { "Authorization": "Bearer TOKEN" },
      "timeoutSeconds": 600
    }
  }
}

• Remote servers must use Streamable HTTP, and url can be used in place of serverUrl.
• The old HTTP+SSE and httpUrl are not supported.
• The default timeout for connecting and listing tools is 300 seconds, and for a tool call it is 180 seconds.
• /mcp opens the management panel.

Skills

Antigravity’s Skills follow the agentskills.io open standard:

• One folder holds one SKILL.md. The frontmatter description is required and name is optional.
• Put workspace skills in .agents/skills/ and global skills in ~/.gemini/config/skills/.
• The CLI also reads ~/.gemini/antigravity-cli/skills/ and ~/.gemini/skills/.
• Each skill becomes a slash command in the CLI. After you change a file, reload it with /skills reload.

What happened to Gemini CLI

Google’s developers blog (2026-05-19) announced that Gemini CLI is moving into Antigravity CLI:

• From 2026-06-18, Gemini CLI and the Gemini Code Assist IDE extensions no longer serve Google AI Pro, Ultra, and free Code Assist individual users.
• Organizations that hold a Gemini Code Assist Standard or Enterprise license are not affected.
• Gemini CLI can still use API keys for paid Gemini or Enterprise Agent Platform.
• Google admits the two are not 1:1 feature-equivalent.
• Concepts that carry over: Agent Skills, Hooks, Subagents, and Extensions (renamed Antigravity plugins).

The Gemini CLI GitHub page now says “Enterprise Code Assist only,” and notes that it is not optimized for Gemini 4 and should not be used to run model benchmarks. You can still install it with npm install -g @google/gemini-cli or brew install gemini-cli. It supports gemini -p and --output-format json, and MCP is configured in ~/.gemini/settings.json. If you are an enterprise user or have an old project, you can keep using it. For new projects, we suggest starting with agy.

Migrating from Gemini CLI

The first time you run agy, it detects your old settings and offers a checklist. You can also run agy plugin import gemini manually.

Jules

Google’s changelog announced on 2026-10-29: Jules will end service on 2026-12-31. Finish any in-progress sessions before 2026-11-11, after which they become read-only.

• You can use Jules Tools (CLI) to export settings, environment variables (including the values), scheduled tasks, linked repos, memories, setup scripts, and all sessions.
• Your code and PRs are already on GitHub.
• The announcement says what Google learned from Jules will be used in Antigravity, but it does not describe a migration path.

So this guide does not treat Jules as a tool you can start using.

Sources: Antigravity getting started, Settings, Rules, MCP, Skills, Migrating from Gemini CLI, Gemini CLI transition to Antigravity CLI announcement, Gemini CLI GitHub, Jules changelog (queried 2026-11-06)

Privacy and Data Controls

How Gemini handles your data depends on which product you use, whether you pay, and whether a given setting is on. This chapter is split into four parts: the consumer app, the API, Workspace, and Antigravity.

Gemini app (consumer)

The core setting is called Keep Activity.

• The default auto-delete is 18 months, which you can change to 3 months, 36 months, or no limit.
• Human review: some chats are reviewed by human reviewers (including trained service providers), and are disconnected from your account before review. Reviewed chats are kept for up to 3 years and are not deleted even if you delete your activity.
• When Keep Activity is off, Gemini cannot connect to Google Workspace, and you cannot see Deep Research history reports.
• Personal Intelligence is off by default and can be switched on or off per app (Chapter 7).

Gemini API and AI Studio

Also: when you offer an API client to users in the European Economic Area, Switzerland, or the UK, you may only use Paid Services.

So the free tier is not suitable for handling customer data or internal documents. Link a billing account.

Google Workspace

The official page says: your company’s data belongs entirely to your company and is not used to train models, and Gemini only retrieves content the user has permission to access. Google also lists ISO 42001 and SOC 1, 2, and 3.

Antigravity

The CLI has an Enable Telemetry option that collects usage data when turned on. It is always off for business accounts.

Recommended settings

1. Open Gemini’s Activity settings page and check the state of Keep Activity and the auto-delete period.
2. Use temporary chat for sensitive content.
3. When you need to connect Workspace or view Deep Research history, know that both depend on Keep Activity.
4. When you process non-public data through the API, use the paid tier.

Sources: Gemini Apps privacy and Activity help, Connect Google Workspace apps, Deep Research help, Gemini API terms, Billing, Google Workspace AI, Antigravity Settings (queried 2026-11-06)

FAQ, Limits, and Prompting Tips

Limits and common pitfalls

Google’s prompting recommendations

The following is compiled from the Gemini API prompting strategies and the Gemini 3 developer guide.

Basics

• State the task directly; specify the length, the format, and what to avoid.
• For complex JSON output, use the structured output feature instead of describing it in text.
• Few-shot: give a few diverse examples with exactly the same format; too many examples cause overfitting.
• Provide reference material and ask it to “answer only from the provided content.”
• Turn on Search grounding for the latest facts and code execution for calculations.

Structure

• Use XML tags or Markdown headings to separate the role, constraints, context, and task. Pick one style.
• For long inputs, put all the context first and the question last, with a transition in between, such as “Based on the information above.”
• Put key constraints and the persona in the system instruction or at the very beginning.

Style and thinking

• From Gemini 3 on, answers are direct and concise by default. If you want a conversational or detailed answer, say so explicitly.
• Models from 2.5 and 3 on reason internally, so you no longer need to ask for “think step by step.” Only add “think carefully” for very hard problems, at the cost of more thinking tokens.
• Thinking level: run your evaluation at all three levels and pick the lowest one that meets your bar.

Agents

• Spell out how to plan, the risks (the difference between reading and writing), when to ask the user, and the retry rules.

Example: applying Google’s recommendations in one prompt

<role>You are a technical documentation editor.</role>

<constraints>
- Answer only from the content of <context>. If you can't find it, say "Not in the material"
- Output a list of at most 5 points, each no longer than 40 words
</constraints>

<context>
(Paste the document content here)
</context>

Based on the information above, summarize the rules related to the refund process.

Sources: Prompting strategies, Gemini 3 developer guide, Gemini subscription plans (US), Gemini subscription plans (Taiwan), Antigravity Models (queried 2026-11-06)

How It Differs from ChatGPT and Claude

This chapter only compares items that have official page backing. ChatGPT and Claude were only verified at the plan-page level and not in their help centers, so there is no price comparison and no judgment about whose model is stronger.

Plan structure

The ChatGPT plan page showed no USD price figures at the time of the query, so they are not in the table above.

Three ways Gemini differs

1. A package tied to your Google account and storage. One subscription includes 400 GB to over 20 TB of storage, YouTube Premium Lite or Premium, Flow credits, and Gemini built into Gmail and Docs. The ChatGPT and Claude personal plan pages do not have this kind of bundling (going by the pages we checked).

2. Development tools are concentrated in one entry point, Antigravity, and you can choose Claude models there. In Antigravity you can choose Claude Opus 5.5 and Sonnet 5.5 (Pro paid plan or above), not just Gemini models (Chapter 4).

3. The API has a free tier, but the data handling terms are different. The free tier is limited to some Flash models, and Argon and 3.1 Pro have no free tier. On the free tier your data is used to improve products and has human review; on the paid tier it does not (Chapter 12).

Data policy

ChatGPT’s data policy was not verified this time, so it is not written up.

What we do not compare

• Model quality: The benchmarks on the Google DeepMind page are published by Google itself, and other vendors do the same. Benchmarks that each vendor publishes cannot be compared with one another. Test with your own tasks.
• Who is cheaper: The three vendors’ usage mechanisms are different. Google uses compute plus a 5-hour and weekly reset; Claude also has 5-hour and weekly limits; ChatGPT has message caps depending on the model. You cannot compare them directly by monthly price.

How to choose

Pick the environment you spend the most time in. If you work in Gmail, Docs, and Drive every day, Gemini’s integration is the smoothest. For which development tool to use, it depends on your project and your team’s current habits. Every vendor’s plans and usage mechanisms keep changing, so before you subscribe, try your own tasks on the free tier or a short-term plan.

Sources: Gemini subscription plans (US), ChatGPT Pricing, Claude Pricing, Antigravity Models, Gemini API terms, DeepMind Gemini models page (queried 2026-11-06)