Learning AI

English

Tiếng Việt

English

Tiếng Việt

Appearance

Gemini (Google) — A multimodal AI assistant that lives inside the Google ecosystem ​

✨

Hands-on — 30 seconds

You're a final-year student. It's 11pm and you owe a literature-review section for your thesis — 20 sources scattered across the web, and reading them by hand would eat your whole night. Open gemini.google.com, turn on Deep Research, and type: "Do deep research on [topic], compare conflicting viewpoints, prioritize 2025–2026 sources, and include a table and citations." → Gemini proposes a plan, you approve it, it browses the web for a few minutes, then returns a sourced report. Export to Google Docs and polish. An all-nighter shrinks to 30 minutes. Real-world payoff: an assistant that can read your Gmail/Drive/Docs, generate images/video, and swallow a 1-million-token document. The Free tier is already enough for most schoolwork. (A one-year free Google AI Pro offer for students ran in late 2025 but registration has now closed — see Section 02 for how to check for a new round.) Less hired help, fewer late nights.

"Gemini isn't just a chatbot — it's a whole family of products: a foundation model, a chat app, an assistant inside Gmail/Docs, and developer tools.It's strongest when you already live inside the Google ecosystem. But it also fabricates citations, and Google just 'retired' the Gemini CLI for individual users — know the details so you don't get caught out."

After this chapter you'll be able to

  • Tell apart the three layers of Gemini (foundation model / chat app / dev tools) so you don't get confused reading the news.
  • Sign in and use the generous Free tier (Deep Research, image generation, Live are all there), and pick a plan that fits your budget.
  • Run Deep Research to produce a report with citations, and combine it with NotebookLM into a powerful research workflow.
  • Use AI Studio (free, no card required) to grab an API key and "vibe-code" an app, then deploy it.
  • Understand the Gemini CLI → Antigravity transition (personal CLI shutdown deadline: Jun 18, 2026) and decide whether to depend on it.
  • Spot & prevent citation fabrication + data leakage on the personal tier — a survival skill when using AI for real work.

Note on the "shelf life" of this information

This reflects understanding as of early June 2026. The Gemini ecosystem changes extremely fast — the first half of 2026 alone brought Gemini 3, 3.1 Pro, I/O 2026, and the Gemini CLI shutdown. Pricing/benchmark figures are tagged with "~" when the source is a third party. Just head straight to gemini.google.com and ai.google.dev to check the latest.

01 · What this tool is & when to use it ​

Gemini is the overall AI product family from Google / Google DeepMind. Unlike ChatGPT (mostly a single app), "Gemini" spans 3 layers that are easy to mix up — get them straight and the news stops being confusing:

Telling apart the 3 layers of "Gemini" (read carefully so you don't confuse them)

  • Foundation models = the Gemini model line underneath, with several tiers: the Pro versions (the most capable: 3 Pro, 3.1 Pro, 3.5 Pro), the Flash / Flash-Lite versions (fast & cheap: 3.5 Flash), and the deep-reasoning Deep Think mode. This is the "engine."
  • End-user apps = the Gemini chat app (web + iOS/Android), the on-phone assistant, and the Workspace integrations. This chapter talks mostly about these.
  • Developer tools = AI Studio, the Gemini API, Vertex AI, and (formerly) the Gemini CLI → now Antigravity.

Gemini 3 (the foundation model) launched Nov 18, 2025, described by Google as its "most intelligent model" — strong reasoning, multimodal (text/image/video/audio/code), with a 1-million-token context window. Gemini 3.1 Pro arrived around Feb 19, 2026. At Google I/O 2026 (around May 19–20) Google announced Gemini 3.5 Flash (now GA — described as fast yet beating 3.1 Pro at coding/agentic work), introduced Gemini Omni, Spark, and Antigravity (an agent-first platform replacing the Gemini CLI). Per sources through mid-2026: Gemini 3.5 Pro is not yet broadly available (it's rolling out gradually), so don't confuse "3.5 Flash" (available) with "3.5 Pro" (not yet). The official app URL: https://gemini.google.com.

🔗 "Shutting down the Gemini CLI" and "launching Antigravity" are the SAME event

The announcement to move from Gemini CLI → Antigravity came at I/O 2026 (~May 19–20, 2026), while the deadline to shut off the Gemini CLI for individual users is Jun 18, 2026 — i.e. about a month after the announcement. Don't misread Jun 18 as "Antigravity's launch day": Antigravity (and the Antigravity CLI) have existed since I/O; Jun 18 is just the day the old CLI stops serving individuals. Details in Section 02.

What Gemini does well (in the app):

Task groupWhat it can doWhere to find it
Multimodal chatQ&A over text/image/video/audio/code; the Free tier runs a Flash-class model (latest generation), paid unlocks the Pro tierEven on Free
Deep ResearchAn agent that browses many web sources, plans, and synthesizes a report with citationsFree (stronger on Pro/Ultra)
Deep ThinkA deep-reasoning mode (many parallel hypotheses); Google announced Gemini 3 Deep Think hitting ~41% on Humanity's Last Exam, ~93.8% GPQA Diamond, and IMO 2025 gold-medal level (figures published by Google)Ultra
Voice & authoringGemini Live (real-time conversation), Canvas (authoring/code workspace), Gems (custom assistants)Free + paid
Image & video generationImages via Nano Banana / Nano Banana Pro, video via Veo 3.1 through FlowBy plan
Personal IntelligenceReads context from Gmail, Drive, Calendar, Maps (with a privacy controls panel)By plan

Two sibling products you must know

  • NotebookLM (notebooklm.google.com) — uses Gemini models, but the workflow is different: you load your own documents as "sources," then ask questions and get answers with citations linking back to the source, generate an Audio Overview (auto-generated podcast), summaries, and slides. It's an extremely powerful personal-document RAG tool — it only answers based on the documents you supply, so it fabricates less than ordinary chat. (Hybrid workflow with Deep Research in Section 03.)
  • Gemini in Google Workspace — integrated into Gmail, Docs, Sheets, Slides, Drive, Meet, Chat, Vids: "Help me write," thread summaries, Sheets formulas, meeting notes. Since ~Jan 2025 it's been bundled free into many Workspace business/enterprise plans.

New at I/O 2026 (rolling out — read with measured expectations)

  • Gemini 3.5 Flash is GA — Google describes it as fast yet beating 3.1 Pro at coding/agentic work; 3.5 Pro is, per sources through mid-2026, rolling out gradually and not yet widely available.
  • Gemini Omni — a model that "creates anything from any input," starting with video, available on Plus/Pro/Ultra. Spark (a 24/7 agent running on 3.5) and Project Genie are for Ultra.
  • Note: these are new I/O 2026 features, some in Beta or limited to the US market — double-check before assuming they're available in your region.

Developer (AI Studio + API): AI Studio (aistudio.google.com) is a free playground — prompt, grab an API key, "vibe-code" an app, then deploy it to Cloud Run; access Gemini/Veo/Imagen/Nano Banana/Gemma. Build mode spins up a full-stack web app or Android (Kotlin/Compose) from natural language.

Use Gemini when: you live inside Google (Gmail/Docs/Drive), need multimodal (image/video/audio), need very long context (1M tokens), or want good local pricing. Think twice when: the task needs absolutely precise citations (legal/medical) — Gemini fabricates citations more than Claude (see below).

Vs. other tools — "when to pick which" ​

No tool "wins on everything." Gemini is strongest at multimodal, long context, and Google integration. The table below synthesizes several third-party comparisons (2026) — benchmarks vary by source, so treat them as directional reference, not absolute numbers:

ToolStrong atCommon paid plan (~USD/month)Pick when
Gemini (Google)Multimodal (image/video/audio), 1M-token context, Workspace integrationPlus ~$8 · Pro ~$20 · Ultra ~$100+You live in Google, need video/images, very long documents, good local pricing
ChatGPT (OpenAI)All-rounder, the widest plugin/integration ecosystemPlus ~$20You need a versatile "all-in-one" assistant
Claude (Anthropic)High-quality code, multi-file reasoning, long/formal documents, safetyPro ~$20Serious programming, long spec-faithful writing, safety-first
PerplexitySearch-first, highest citation accuracyPro ~$20Lookups needing trustworthy sources, fast research with citations
Grok (xAI)Direct access to the X feed, fastest speed~$30 (varies by X plan)Tracking breaking news / real-time social media

Reference USD prices (global, per multiple sources through mid-2026), may vary by promotion & region — check the localized price shown on the official site for your country.

Quick "who's strong at what" summary

  • Multimodal + long context + work integration → Gemini (plus Veo/Nano Banana).
  • Code quality & multi-file reasoning → Claude (one source reports SWE-bench Verified ~87.6% vs Gemini 3.1 Pro ~80.6% — third-party figure, read with measured expectations).
  • Citation accuracy / real-time search → Perplexity (lowest citation-hallucination rate).
  • Speed + social-media data → Grok (direct X access).
  • Running two tools in parallel is totally normal — Gemini for thinking/research, Claude for coding, for instance.

CLI coding specifically (Antigravity vs Claude Code vs Codex) ​

If you plan to code with a CLI agent, this is the head-to-head comparison most worth weighing. Antigravity's biggest point of contention is that its free quota dropped drastically versus the old Gemini CLI:

AxisAntigravity CLI (agy)Claude CodeCodex CLI
Open source?Closed-sourceClosed-sourceOpen source
LanguageGo (binary, no Node needed)Node.jsRust
Free quota/dayper community reports through mid-2026: ~20 runs/day (the old Gemini CLI was ~1,000/day → a ~98% cut)By Claude plan (Free limited; Pro/Max more)By ChatGPT plan
Agent featuresKeeps Agent Skills, Hooks, Subagents, Extensions; supports MCPSkills, Hooks, Subagents, MCPTool use, MCP
MaturityNew (~I/O 2026), community complains about insufficient feature parityMature, widely usedMature

Antigravity's free quota of ~20 runs/day is a real pain point

Per community reports through mid-2026: Antigravity CLI's free quota fell to about ~20 runs/day — just a few tasks and it's gone, a far cry from the old Gemini CLI (~1,000/day). If you code a lot and don't want to pay, consider Claude Code / Codex as a fallback, or upgrade to Gemini Code Assist Standard for a higher quota. This number changes often by round — recheck before depending on it.

When NOT to use Gemini (real limits)

Gemini is good at many things, but not every job suits it. Avoid it — or always keep a human in the loop — in these cases:

  • Legal / medical content needing absolutely precise citations — one source reports Gemini "hallucinating" case-law citations ~18% of the time (vs Claude ~3%). That figure is from a third-party comparison, not formal research — but it's enough to not trust it blindly for legal/medical content; always verify.
  • Sensitive / confidential data on the personal tier — because of human review + long retention (see Section 04). Workspace enterprise is different.
  • Production pipelines needing an open-source, long-term-stable CLI — the Gemini CLI was just "retired" for individuals (Jun 18, 2026), and Antigravity is closed-source and lacked feature parity at launch → if you need open-source/stability, consider Claude Code/Codex or wait for Antigravity to mature.
  • Large codebases needing top reliability for multi-file edits — many benchmarks give Claude the edge.
  • Tasks needing real-time social-media data — Grok (X access) fits better.
  • Needing absolute API stability right when a model just launched — the Preview phase often hits 503 "overloaded" errors (see Section 04).
  • Recognizing specialized or regional accents in Gemini Live / audio — spoken-language recognition quality (especially terminology and regional accents) is often weaker than for English; test first with a sample clip before relying on it for important work.
  • Needing stable / reproducible results — the model changes versions constantly (3 → 3.1 → 3.5 in just 6 months), and output can drift between generations. Don't hard-code a rigid dependency on one version in a product; pin a specific version via the API if you need stability.

02 · Sign-up & access — pricing & access ​

Available worldwide? — Yes. ​

Google officially sells all 3 plans in most regions (no VPN needed). The Free tier requires no card. Prices are shown in your local currency on the official subscriptions page.

The one-year free Pro student offer — REGISTRATION HAS CLOSED

In late 2025 Google ran an offer giving students one year of Google AI Pro free (at the time "powered by Gemini 2.5 Pro" + 2TB), with registration via SheerID verification. Per sources through mid-2026: the registration window opened around Oct 8, 2025 → Dec 9, 2025 in some regions and this 12-month round has closed globally (regional deadlines extended to ~March 2026). As of now (mid-2026) registration is closed — don't assume it's still available. Google often reopens these in rounds, so check gemini.google/students to see whether a new one is live.

30-second sign-in (the app — nothing to install) ​

text
1. Open https://gemini.google.com (or download the iOS / Android app).
2. Sign in with your Google account.
3. Go straight to the chat screen → type and you're ready to go (Free tier).

Plans & pricing (local prices come from gemini.google/subscriptions, shown in your currency) ​

PlanPricingKey contents
Free03.5 Flash + several tiers of 3.1 Pro; image generation, Deep Research, Live, Canvas, Gems; 15 GB
AI Plus~$8/month (often 50% off for the first 6 months)2× quota, video generation, 200 Flow credits, advanced NotebookLM, 200 GB
AI Pro~$20/month4× quota, "3.1 Pro" + advanced features, 1,000 Flow credits, 5 TB, YouTube Premium Lite
AI Ultra~$100+/month (with higher tiers)Deep Think (early, US/English), 10,000–25,000 Flow credits, 20+ TB

Notes on pricing (read before opening your wallet)

  • Reference USD prices (global, per multiple sources): Plus ~$7.99, Pro ~$19.99, Ultra ~$99.99–$249.99 (I/O 2026 lowered the top tier). There's slight variation between sources → the number shown on your country's official page is the most reliable.
  • The Free tier is quite generous — it already has Deep Research, image generation, and Live. Enough for most learners. Don't rush to pay.
  • Students: the one-year free Pro offer has closed registration (see the box above) — if Google opens a new round, jump on it; otherwise just start with Free.

Payment ​

Through Google Play / Google One with an international card (Visa/Mastercard); some local cards/wallets work via Google Play depending on region.

Note for Vietnam / SEA readers — local payment methods

The exact list of local payment methods (which wallets/cards) varies over time and region, and reliable sourcing is thin. Check directly on Google One at purchase time. If a local card is declined, try an international Visa/Mastercard or the channel Google One suggests for your region.

Developer — AI Studio is free, no card required ​

text
1. Open https://aistudio.google.com → sign in with Google.
2. Grab an API key (there's a free tier) or open the Build tab to "vibe-code" an app.
3. No credit card needed to start.

Notable change: the Pro model leaves the API free tier

  • The API free tier still has Flash / Flash-Lite free, but as of Apr 1, 2026 the Pro model is removed from the free tier → to use Pro via the API you must pay.
  • The free tier is capped at ~5 requests/minute — fine for experimenting, too little for a multi-user app → upgrade the tier or throttle client-side.
  • Reference API pricing (third-party): Gemini 3.1 Pro ~$2.00 input / $12.00 output per 1M tokens (≤200K); above 200K it rises to ~$4/$18. Flash-Lite ~$0.10–$0.40/1M. The Batch API is roughly half the price.

Free-tier AI Studio / Gemini API data MAY be used to improve the model

A survival point for devs and prototypers: per Google policy through mid-2026, prompts/content sent through the free tier (AI Studio / unpaid Gemini API) may be used to improve the product (including human review). The paid tier does NOT use your data for training. → Don't prototype with customer data / real data on the free tier. Before committing, check the current policy at ai.google.dev/gemini-api/terms.

Installing a CLI coder — Antigravity CLI (replaces the Gemini CLI) ​

Google is shutting down the Gemini CLI (open-source) for Free/AI Pro/Ultra users as of Jun 18, 2026, replacing it with the Antigravity CLI (closed-source, written in Go, command agy). Here's the standard way to install from now on:

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

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

# Run (binary named 'agy', NO Node.js needed)
agy

The first time you run agy it opens the browser to sign in with Google; over SSH/headless it prints a URL + one-time code. Antigravity keeps Agent Skills, Hooks, Subagents, Extensions (packaged as plugins).

Installing the old Gemini CLI (only relevant for Standard/Enterprise orgs now)

For individuals, the Gemini CLI stops serving as of Jun 18, 2026 — this section only still works if you're a Gemini Code Assist Standard/Enterprise customer (paid API key) or an org using it through Google Cloud.

bash
npx @google/gemini-cli            # run without installing
npm install -g @google/gemini-cli  # install globally (needs Node 18+, 20+ recommended)

IMPORTANT TURNING POINT: Gemini CLI → Antigravity (read carefully if you code via CLI)

Google is moving from Gemini CLI (open-source) → Antigravity CLI (closed-source) for individual users as of Jun 18, 2026. The Gemini CLI is kept only for Standard/Enterprise customers. The community reacted strongly because:

  • It goes from open to closed (losing transparency / the ability to patch it yourself).
  • It lacked feature parity at launch (The Register, FOSS Force, Hacker News all noted this).

→ If your pipeline needs an open-source, long-term-stable CLI, consider Claude Code / Codex, or wait for Antigravity to mature. Don't build a critical workflow with a hard dependency on a CLI that's mid-transition.

03 · Hands-on workflows — step by step (with real prompts) ​

Here are 4 workflows that go "from start to finished." Each step has a way to self-check (verify) so you know you did it right.

A) Deep Research in the app — a report with citations ​

Step 1 — Turn on Deep Research. Open gemini.google.com → select Deep Research (or enable it from the tools menu). → Verify: you see Deep Research mode selected in the toolbar.

Step 2 — Type a clear prompt.

text
Do deep research on [topic]. Outline the sub-topics, compare conflicting
viewpoints, prioritize 2025–2026 sources, and include a comparison table
and a list of citations. If any source is uncertain, say so instead of guessing.

Step 3 — Review the plan. Gemini proposes a plan (the sub-topics it will search) → you edit/approve it → it browses the web for a few minutes → returns a report. → Verify: the plan matches your intent; the report has clickable source links, not 404 links.

Step 4 — Export to Docs. Click to export to Google Docs to edit and save. → Verify: opening the Doc shows the content + full citations.

B) The Deep Research + NotebookLM hybrid workflow (very powerful for study/research) ​

This is the "sweet spot" combo of the Google ecosystem — pairing broad web search (Deep Research) with personal-document RAG that cites back to the source (NotebookLM):

text
1. Gemini Deep Research: gather web info on the topic → copy/export the report.
2. Go to notebooklm.google.com → create a new notebook.
3. Add that report as a Source + add your own original PDFs/documents.
4. Ask questions with citations linking BACK to the source; generate an
   Audio Overview (podcast); cluster insights into a framework.

→ Verify: every answer in NotebookLM links back to the right passage in the source you loaded (this is the difference — NotebookLM only answers based on your sources, so it fabricates less than ordinary chat).

C) Coding with the Antigravity CLI ​

bash
agy                      # open the agent TUI in the project folder
# in the session: describe the task in your language, the agent reads/edits
# multiple files, calls tools, and keeps the session history

Configure project context — create a file describing the project (inheriting the context file concept like the Gemini CLI's GEMINI.md): put the project description, code conventions, and build commands so the agent reads the right context:

text
# Project context file (example contents)
Project: an order-management web app (React + Supabase).
Conventions: TypeScript, no 'any'; components live in src/components.
Build command: npm run build. Test command: npm run test.
DO NOT edit files in the /legacy folder.

Antigravity's context-file name & MCP paths DIFFER from the old Gemini CLI

The old Gemini CLI used GEMINI.md for context and ~/.gemini/settings.json for MCP. Antigravity changed the convention — the context-file name, format, and MCP config paths are not the same as the old CLI. Don't follow the convention of a retired tool: look up the exact file name & paths at antigravity.google/docs/cli-getting-started before creating anything.

→ Verify: run agy in the repo, and the agent reads the context file and follows the conventions you wrote (e.g. doesn't touch /legacy).

D) AI Studio Build mode — build an app with no infrastructure ​

text
1. aistudio.google.com → Build tab → describe the app in your language.
2. The agent generates code (auto-creating placeholder images with Nano Banana)
   and shows a preview.
3. Deploy to Cloud Run RIGHT inside AI Studio,
   or grab an API key to embed in your own app.

→ Verify: the preview runs as described; if you deploy, the Cloud Run URL opens.

Git tip before letting the agent edit files

Before running agy (or any agent that auto-edits files): commit a clean state to git first, work in a clean repo, review each diff, and split tasks small. There are reports of the CLI deleting/editing code by mistake in long sessions (see Sections 04/06) — committing first is the cheapest safety net.

04 · Handy tips & common mistakes ​

🟢 Money-making tips ​

7 tips to use Gemini like a pro

  1. If you live in Google, turn on Personal Intelligence — let Gemini read Gmail/Drive/Calendar so it understands your work context (but see the privacy box below about sensitive data).
  2. Research → Deep Research; your documents → NotebookLM. Don't make ordinary chat carry both — NotebookLM cites back far more reliably.
  3. The 1M-token context is a real advantage — paste an entire long document / many files at once instead of chopping them up.
  4. Let the AI say "I'm not sure": add "if there's no source, say you're not sure" → clearly reduces fabrication, especially with citations.
  5. Hit a 503 "overloaded"? Switch to Flash. The Flash model recovers faster than Pro when servers are overloaded (see FAQ).
  6. Letting an agent (CLI) edit files → always in a clean git repo + review the diff. Long sessions risk deleting things by mistake.
  7. Watch for student offers: activate a free year of Pro before thinking about paying, if a round is open; AI Studio is also free for devs.

🔴 Errors & pitfalls (read carefully — this part saves you) ​

🚨 Citation hallucination — Gemini's #1 pitfall

Gemini can very confidently cite sources that don't exist, especially in legal/medical content. One third-party source reports a citation-hallucination rate of ~18% (vs Claude ~3%) — that figure isn't formal research, but it's enough to:

  • Not trust it blindly for legal/medical content.
  • ALWAYS verify every citation/figure before putting it in an official document.
  • Prefer NotebookLM (cites back to the sources you loaded) or Perplexity when you need high source accuracy.

Other traps to remember

  • The CLI deleting/editing code by mistake in long sessions ("state degradation") → always work in a clean git repo, review each diff, split tasks small.
  • The Pro model vanishing from the API free tier (as of Apr 1, 2026) → only Flash/Flash-Lite remain free; switch models or pay.
  • The Gemini CLI ceasing to serve individuals (Jun 18, 2026) → move to agy (Antigravity) or a Standard/Enterprise license.
  • The API free tier's 5 req/min is too little for a shared app → upgrade the paid tier or throttle client-side.
  • I/O 2026 features (Gemini Omni/3.5/Spark) are still rolling out, partly US-only — don't assume they're available in your region.

Privacy & data — read carefully if using it for work

This part really matters if you use Gemini for company work. Information per Google's documentation through early 2026:

(a) The Gemini app (personal account) — where does the data go?

  • By default it saves conversations (Gemini Apps Activity), with default retention of 18 months (changeable to 3 or 36 months).
  • Google samples some chats for human review (including service-provider reviewers) to improve the model; a conversation a reviewer touches can be kept up to 3 years — and you shouldn't paste sensitive data into it.

(b) How to turn off training / human review (personal plan):

  • Go to myactivity.google.com/product/gemini → turn off Keep Activity; or myaccount.google.com → Data & Privacy → Gemini Apps Activity → Turn off.
  • Once off, new chats aren't sent for review / aren't used for training.

(c) Gemini in Workspace (enterprise) — completely different from the personal tier:

  • Prompts and content stay within the organization, aren't shared externally, and aren't used to train public models; enterprise controls apply (data region, DLP).

(c2) AI Studio / Gemini API: per policy through mid-2026 — on the free tier prompts may be used to improve the model; on the paid tier they are not. → Don't prototype with customer data on the free tier (see Section 02). Check ai.google.dev/gemini-api/terms before committing.

(d) NEVER paste the following into the personal tier:

  • National ID numbers, bank card numbers, passwords, OTPs.
  • Contracts/NDAs, confidential documents, proprietary source code.
  • Customers' personal data (names, phone numbers, addresses, records) — under data-protection law (e.g. GDPR, or your local privacy regulation), this can constitute a violation.

(e) When letting an agent (CLI) edit files automatically: there's a risk of deleting/editing by mistake → commit to git first, review the diff.

FAQ & common errors (click to open)

503 "model is overloaded" / "Deadline Expired" — what now? This is Google's servers overloaded, NOT a quota/billing error on your end. How to handle it: (1) don't increase the timeout; (2) use exponential backoff retry; (3) wait 5–30 minutes (≈70% recover within 60 minutes); (4) fall back to Gemini 2.5/3 Flash (recovers fast, 5–15 minutes). Common right after a Preview model launches (e.g. the Feb 19 & Feb 26, 2026 rounds).

429 RESOURCE_EXHAUSTED — is it the same as 503?Completely different. 429 is your personal quota / rate limit → reduce frequency, upgrade the tier, or use the Batch API. Don't confuse 429 (your error) with 503 (Google's server error).

The API free tier is only 5 req/min — what about my multi-user app? Too little for a shared app → upgrade to the paid tier or throttle client-side (queue requests). Consider the Batch API if you run in batches.

The CLI deleted/edited my code by mistake? This has been reported in long sessions. Prevent it: always work in a clean git repo, review each diff, split tasks small, and commit often.

I'm an individual and the Gemini CLI says it's ceasing service? Correct — as of Jun 18, 2026 the Gemini CLI stops serving individuals. Move to agy (Antigravity) or use a Standard/Enterprise license.

Gemini answers in English even though I asked in another language? Add an explicit instruction: "Always answer in [language]." or set it in Gems/Saved info so you don't have to repeat it.

The Pro model vanished in AI Studio? As of Apr 1, 2026 the Pro model was removed from the API free tier — only Flash/Flash-Lite remain free. Switch models or pay.

The Antigravity CLI (agy) reports out of quota after ~20 runs/day? This is the community's most common complaint — per reports through mid-2026, Antigravity's free quota is only ~20 runs/day (the old Gemini CLI ~1,000/day). How to handle it: check the remaining quota in-session; if you code a lot, upgrade Gemini Code Assist Standard for a higher quota, or use Claude Code / Codex as a fallback. The quota number changes often by round — recheck at antigravity.google/docs.

After installing Antigravity, typing agy says "command not found"? The installer drops the binary into ~/.local/bin, which may not yet be in your PATH. The fix: add ~/.local/bin to PATH (e.g. in ~/.zshrc / ~/.bashrc) then reopen the terminal; or call it by full path ~/.local/bin/agy to test first.

SheerID can't verify my university (when registering for the student offer)? A classic error while the offer is open: SheerID sometimes doesn't have your school's name in its list. The common workaround: pick "school not in the list" then upload proof (student card / .edu email), or retry with the school's official English name. Note: as of mid-2026 registration has closed (see Section 02) — this only helps if Google reopens a new round.

05 · Exercises / mini-projects ​

Actually do 2–3 of the exercises below to turn "I understand" into "I can do it." Each has clear completion criteria.

Exercise 1 — Verified Deep Research (basic) ​

Goal: run a Deep Research report and build the reflex of verifying sources.

  1. Go to gemini.google.com → turn on Deep Research.
  2. Use the prompt:
text
Do deep research on "electric vehicle trends in Southeast Asia 2025–2026."
Compare conflicting viewpoints, include a data table and a list of citations.
If any figure lacks a solid source, mark it "uncertain" instead of guessing.
  1. Open any 3 source links in the report and check whether they're real (not 404) and actually match what Gemini cited.

Done when: you have a report + confirmed at least 3 sources are real and matching (or caught a fabricated/wrong source — an even more valuable lesson).

Exercise 2 — Personal-document RAG with NotebookLM (important) ​

Goal: experience citing-back — the strongest fabrication-reducer in the Google ecosystem.

  1. Go to notebooklm.google.com → create a notebook → load 1–2 of your own PDFs (study materials, a report) as Sources.
  2. Ask:
text
Summarize this document into 5 key points. For each point, cite back to
the exact passage in the source. If information isn't in the document,
clearly state "not in the source."
  1. Click each citation-back and check it jumps to the right passage in the original file.
  2. (Optional) Generate an Audio Overview to hear the auto-generated podcast.

Done when: every point links back to the right source passage; you clearly see NotebookLM sticking to the document rather than fabricating extra.

Exercise 3 — Vibe-code a small app in AI Studio (lightly advanced) ​

No card required

AI Studio is free, no credit card needed to start.

Goal: build a working app from a natural-language description.

  1. Go to aistudio.google.com → Build tab.
  2. Describe the app, for example:
text
Create a "split-the-bill" web app: enter the total bill and the number of
people, show the amount per person, with a round-up button. Clean,
mobile-friendly UI.
  1. Check the preview, refine with a few more sentences if needed.
  2. (Optional) Deploy to Cloud Run right inside AI Studio.

Done when: the preview app works correctly (enter numbers → get the right split). Getting a Cloud Run URL deployed is a bonus.

Exercise 4 — Coding with the Antigravity CLI in a clean repo (advanced) ​

Quota note

Antigravity's free quota is ~20 runs/day (per community reports through mid-2026) — enough for this small exercise, but don't waste it. Always commit a clean state to git first before letting the agent edit files.

Goal: experience a CLI agent editing real files + the safety reflex (git + review the diff).

  1. Install the Antigravity CLI (see Section 02). If typing agy says "command not found" → add ~/.local/bin to PATH.
  2. Go into a small repo (or create one), git init + commit a clean state.
  3. Create a project context file (look up the correct file name at antigravity.google/docs/cli-getting-started), with 1–2 conventions, e.g. "DO NOT edit the /legacy folder."
  4. Run agy, give it a small task in plain language (e.g. "add a function that totals the bill and write a test for it").
  5. Review each diff before accepting; run git diff to check whether the agent touched any out-of-scope files.

Done when: the agent finishes the task following the conventions (doesn't touch /legacy), you can review the diff, and you know how to git restore if it edits something by mistake.

06 · Case studies & real use-cases (from the community) ​

This section gathers real examples from blogs, official Google announcements, and aggregated community discussion (Hacker News, Medium) through early 2026. The point: to show you how Gemini runs in the real world — both when it shines and when it becomes a trap.

Read carefully about source reliability

Some of the content below is personal / second-hand experience — i.e. dev blogs, Hacker News posts, or Google's own vendor claims. So:

  • Statements like "convenient but it misses things" are the author's subjective assessment → read with a balanced attitude.
  • Figures Google publishes itself (e.g. Deep Think winning IMO gold) are vendor claims — weigh accordingly.
  • Whatever's certain (official Google / press) is marked as such; whatever's just "per a personal blog" is also marked.

CS1 — Materials-science literature review: convenient, but it misses deep papers ​

  • Context: A researcher (computational-materials PhD background) tested Deep Research to do a literature review on "generative models for inorganic crystal structures."
  • What they did: Cross-checked the Deep Research results against 27 papers in their personal library.
  • Result / lesson: Convenient, but each deep-research tool has a different "taste" in picking papers, and often misses deep / paywalled material. → Deep Research is good for sketching a framework & finding directions, but can't replace reading the original papers yourself for serious research.
  • Source: Xiangyu Yin's blog (2026) — a researcher's personal experience.

🏛️ CS2 — STOC 2026 used Gemini to auto-generate feedback for submissions ​

  • Context: STOC 2026 (a top theoretical computer science conference) used Gemini to generate automated feedback for submissions.
  • Result: There's a testimonial from Prof. Shuchi Chawla praising results that exceeded expectations.
  • Lesson: Gemini is strong enough to support serious academic processes (screening/feedback), not just casual chat. But this is a context with expert oversight — not a replacement for human reviewers.
  • Source (official): research.google blog (2026).

CS3 — Master's thesis on 5G Anomaly Detection: build a "skeleton," then edit right in Docs ​

  • Context: A student doing a master's thesis on 5G Anomaly Detection.
  • What they did: Used Gemini to build a ~1,000-word "skeleton" for the literature-review section, then edited it right in Google Docs.
  • Result / lesson: This is a practical sweet spot of Gemini for thesis writers — frame the structure fast + refine in Docs (Workspace integration), instead of writing from a blank page. You still have to rewrite/verify the academic content yourself.
  • Source: UCStrategies / aggregated sources (2026).

CS4 — NotebookLM for Quarterly Business Reviews & sales battlecards ​

  • Context: Enterprise teams used NotebookLM for Quarterly Business Reviews (QBR) and sales.
  • What they did: Loaded strategy documents → got a storyline for the deck, talking points for leadership, and a Q&A where every answer links back to the source. The sales team turned product briefs / pricing / win-loss into a battlecard.
  • Result / lesson: NotebookLM shines at internal-document RAG with citations — exactly the work that needs reliability (when leadership asks "where's the source?", you click to show it instantly). This pattern applies to any team with its own document repository.
  • Source: UCStrategies; Medium @nitinfab (2026).

CS5 — Gemini 3 Hackathon ($100k): build a real product under a time constraint ​

  • Context: The Gemini 3 Hackathon ($100k, Dec 2025–Feb 2026).
  • What they did: Developer Nisrine Amimi recounts the experience of building a real product under a time constraint with Gemini 3.
  • Result / lesson: She describes it as "completely different from just talking about AI" — i.e. the real value comes from getting hands-on and shipping a product under a deadline, not from reading demos. The lesson for learners: you absorb AI far better by doing real projects.
  • Source: Medium / Google Cloud Community — personal experience.

CS6 — Vibe-coding with Gemini 3 in the CLI: strong "one-shot" but the CLI deleted code by mistake (a real trap)

  • Context: A "5 things to try" piece on vibe-coding with Gemini 3 Pro in the Gemini CLI, discussed on Hacker News.
  • What happened: The community noted the model is strong at "one-shot" (finishing in a single pass), but there was an incident of the CLI deleting code by mistake in a long session.
  • Lesson: This is a classic trust-boundary warning — when letting an agent auto-edit files, always commit to git first, review the diff, and split tasks small. "One-shot" power doesn't offset the risk of losing code if you have no safety net.
  • Source: Hacker News (2026) — aggregated community discussion.

⚖️ A dissenting view (for balance)

A former Google employee once called Gemini the "most frustrating" model they'd used for dev work (paraphrased, no @handle attached). We raise this not to bash it — but so you see that not every experience is positive. Real-world experience depends heavily on the task type: Gemini shines at research/multimodal/long-context, but for multi-file coding many people still find Claude has the edge.

Notable sources (title + context) ​

These are accessible sources — official Google, personal blogs, and community discussion:

Official sources with links (trustworthy):

Illustrative examples (not citable sources — just retold so you get the picture):

  • The materials-science literature-review experience (a researcher's personal blog) — illustrative, no verifiable URL attached.
  • STOC 2026 using Gemini for submission feedback (research.google blog) — official Google, search research.google if you need to verify.
  • NotebookLM for enterprise QBR & battlecards (aggregated Medium/UCStrategies) — illustrative, no verifiable URL attached.

07 · Summary & Official sources ​

5 things to take away

  1. Gemini = a whole product family (foundation model + chat app + Workspace + dev tools), strongest when you live inside Google.
  2. Multimodal + 1M-token context + Deep Research/NotebookLM is the sweet spot — research, image/video, long documents.
  3. Anti-fabrication = let it say "I'm not sure" + prefer NotebookLM (citing back) + ALWAYS verify (Gemini fabricates citations more than Claude).
  4. It's available legally worldwide; the Free tier is generous enough for most schoolwork. (The one-year free Pro student offer has closed registration — check gemini.google/students for a new round.)
  5. The CLI turning point: the Gemini CLI stops serving individuals Jun 18, 2026 → use agy (Antigravity), but the free quota is only ~20 runs/day. When letting the agent edit files, commit to git + review the diff.

These are the official pages for checking the latest info yourself — always trust these links over third-party roundups:

Confidence notes (research through early June 2026)
  • Certain (official Google sources): localized prices on the subscriptions page, the Gemini 3 launch date (Nov 18, 2025), the Gemini CLI shutdown on Jun 18, 2026, the Antigravity install commands, the privacy/retention policy of 18 months, AI Studio being free.
  • Fairly certain (reputable press): the ~$5 entry plan. The one-year free Pro student offer did genuinely exist but registration closed (windows ~Oct 8–Dec 9, 2025 in some regions) — no longer in effect at present.
  • Reference/third-party (tagged "~"/hedged): all comparison benchmarks (SWE-bench Claude ~87.6% vs Gemini 3.1 Pro ~80.6%, citation-hallucination rate 18% vs 3%, the Deep Think scores), per-token API pricing, Antigravity's free quota ~20/day vs the old Gemini CLI ~1,000/day, the new Antigravity config paths.
  • Thin sourcing / links to check at build time: the list of local payment methods; the names/details of I/O 2026 models (3.5 Flash is GA, 3.5 Pro/Omni/Spark still rolling out, partly US-only); the URLs gemini.google/subscriptions/ and support.google.com/gemini/answer/13594961 (the support article ID may change — open it to check before trusting it).

Figures (prices, models, features) may have changed — always recheck at gemini.google.com and ai.google.dev.