Google AI

Best for: Picking the right consumer subscription. Google consolidated its old "Gemini Advanced" / "Google One AI Premium" plans into a cleaner Free / Pro / Ultra structure.

• Free ($0/mo) — limited Gemini 3.1 Pro queries, basic Deep Research, NotebookLM with caps, free Veo 3.1 clip generation in Vids (rate-limited)
• Google AI Pro ($19.99/mo or $199.99/yr) — higher Gemini 3.1 Pro and Deep Research limits, NotebookLM Plus, advanced Vids, Imagen and Veo with bigger quotas, 2 TB storage
• Google AI Ultra ($249.99/mo) — Deep Think, Gemini Agent, Cinematic Video Overviews, up to 1,000 Veo videos/month, 30 TB storage, $100/mo Google Cloud credits

Limitations: The Pro→Ultra jump is steep ($230/mo). Most individual users land on Pro; Ultra is built for power video creators, agentic workflows, and Deep Think users. Workspace customers get Gemini features bundled separately and should not pay for AI Pro on top.

Best for: Teams already on Google Workspace. Gemini is now bundled directly into all paid Workspace plans rather than sold as a separate add-on.

• Business Starter ($8.40/user/mo) — Gemini side panel in Gmail/Docs/Sheets/Slides/Meet, basic AI features, 30 GB pooled storage
• Business Standard ($16.80/user/mo) — everything above plus 2 TB pooled, Meet recording with AI notes
• Business Plus ($26.40/user/mo) — Vault retention, eDiscovery, advanced endpoint management
• Enterprise Standard / Plus (custom) — typically $30–$60/user/mo, S/MIME encryption, DLP, advanced security
• Workspace data stays inside the customer boundary — not used for training. SOC 2/3, ISO 27001/27017/27018, HIPAA BAA available

Limitations: AI features are bundled at no extra cost on Business plans through at least May 31, 2026; Google has signaled an AI add-on may return after that. Enterprise pricing is opaque and quote-based. Workspace AI is not a substitute for the consumer Gemini app's Deep Think or Gemini Agent.

Best for: Free, browser-based prototyping environment for the Gemini API. The fastest way to get a Gemini API key, test prompts, and export working code.

• Free access to Gemini 3 Flash and 3.1 Flash-Lite for prototyping; multimodal playground for text, image, audio, and video inputs
• Test structured output, function calling, JSON mode, and grounding before committing to production
• One-click "Get code" exports a working snippet in Python, JavaScript, or curl
• Issues API keys directly — no separate Google Cloud setup required for the developer tier
• Gemma 3/4 model testing and Firebase integration for full-stack prototypes

Limitations: Free tier rate limits roughly 5–15 RPM and 100–1,000 RPD per model with a 250K tokens/min ceiling. Gemini 3.1 Pro is paid-only as of April 2026 (no free-tier access in API or AI Studio). Free-tier prompts may be used to improve products — switch to paid for production. Not for scale.

Best for: Production-grade enterprise AI on Google Cloud. Vertex AI hosts Gemini, Imagen, Veo, Lyria, Chirp, plus Anthropic Claude and open Gemma/Llama models — all under GCP enterprise governance. Google has begun rebranding Vertex AI as the "Gemini Enterprise Agent Platform" in 2026.

• Model Garden — first-party Google models alongside Anthropic Claude, Mistral, and open-weight Gemma/Llama, all behind one billing surface
• Agent Builder / Agent Engine — managed runtime for agents with sessions, memory, tool governance, and code execution
• Vertex AI Search — grounded enterprise RAG with citations; $1.50 per 1,000 queries (Standard) or $4.00 (Enterprise with generative answers)
• Evaluation Service, Pipelines, Training, and full MLOps tooling
• VPC-SC, CMEK, IAM, audit logging — full GCP enterprise compliance posture
• Pay-as-you-go: token-based Gemini pricing matches API rates; Agent Engine code interpreter $0.0864/vCPU-hr + $0.0090/GB-hr

Limitations: Billing complexity (token + compute + storage + service add-ons) makes cost forecasting hard. Steeper learning curve than AI Studio. The Vertex AI ↔ "Gemini Enterprise Agent Platform" rebrand is partial — both names appear in current Google docs.

Best for: Building Gemini into your own application. Two routes: Google AI for Developers (AI Studio API keys, fastest path) or Vertex AI on GCP (enterprise governance, regional endpoints).

• Gemini 3.1 Pro: $2.00 input / $12.00 output per 1M tokens; 1M-token context (64K output cap); doubles to $4 / $18 above 200K input tokens; paid-only
• Gemini 3 Flash: $0.50 / $3.00 per 1M; 1M context; free tier available — the everyday workhorse
• Gemini 3.1 Flash-Lite: $0.25 / $1.50 per 1M; 1M context (65K output); free tier available — cheapest, ideal for high-volume agent loops
• Gemini 3.5 Flash (released May 19, 2026): $1.50 input / $9.00 output per 1M tokens — beats Gemini 3.1 Pro on coding at ~25% lower cost
• Multimodal input (text, images, audio, video, PDFs); structured output / JSON mode; function calling; code execution; grounded generation with Google Search
• Context caching, batch API, and streaming responses all supported

Limitations: Free tier prompts may be used to improve products — the paid tier is not. Regional availability varies. Gemini 3.1 Pro is excluded from the free tier. Reasoning patterns and instruction-following differ from OpenAI and Anthropic models — expect some prompt-rewriting when porting.

Best for: Building software with Gemini 3.1 Pro at the center of an agent-first development environment. Launched November 18, 2025 alongside Gemini 3, Antigravity is Google's standalone IDE (a VS Code fork) built around a "Manager Surface" for orchestrating multiple parallel coding agents.

• Models: Gemini 3 / 3.1 Pro by default, with Claude Sonnet 4.6, Claude Opus 4.6/4.7, and GPT-OSS 120B available
• Manager Surface — spawn, observe, and orchestrate multiple async agents across workspaces
• Editor view with traditional tab completion, inline commands, and VS Code extension compatibility (via Open VSX)
• Built-in browser preview — agents drive a real browser and produce verifiable artifacts (screenshots, traces)
• MCP support added in early 2026 for GitHub, databases, APIs, and custom tools
• Standalone app for macOS, Windows, and Linux — no extension form
• Pricing: Free public preview (rate-limited), AI Pro $20/mo, AI Ultra $249.99/mo (shares the Google AI plans)

Limitations: Public preview status — rapid changes and stability hiccups. Users have reported rate-limit tightening on free tier and opaque credit pricing on paid tiers. Standalone IDE only (no JetBrains, Visual Studio, or Vim integration). Open VSX marketplace means some proprietary VS Code extensions are missing. Less mature as a general-purpose editor than Cursor or Windsurf.

Best for: Source-grounded research. You upload PDFs, Docs, slides, websites, or YouTube videos; NotebookLM answers questions strictly from those sources with citations and generates rich multimedia briefings.

• Audio Overviews — podcast-style two-host conversations summarizing your sources
• Video Overviews — whiteboard, kawaii, watercolor, and classic visual styles
• Cinematic Video Overviews (launched March 2026) — animated narrated explainers, Ultra-only
• Mind Maps, Slide Decks, Infographics, Quizzes, and Flashcards — all generated from your sources
• Source-grounded chat with inline citations — answers cannot wander beyond your uploaded materials

Limitations: Free tier capped at 3 Audio Overviews/day with limited notebook and source counts. Cinematic Video Overviews require Google AI Ultra ($249.99/mo). Not real-time web-connected — you must supply the sources. Best for research/study, not general chat.

Best for: Quick AI-assisted answers without leaving Google Search. Two surfaces: AI Overviews (above search results) and AI Mode (a dedicated tab for multi-step research conversations).

• Gemini 3 powers AI Overviews globally as of 2026
• AI Mode uses "query fan-out" — spawns parallel sub-searches for multimodal and comparative questions
• Deep Search inside AI Mode runs hundreds of searches and returns a fully cited expert report
• Multimodal inputs: voice, image, and Lens (camera-driven search)
• Personal Intelligence (opt-in) connects Gmail and Photos for personalized travel/shopping responses

Limitations: Free, ad-supported. Some AI Mode features are region-gated. Personal Intelligence requires explicit consent. Publishers continue to express concern about reduced referral traffic. Not a replacement for the Gemini app on serious research tasks — AI Mode is faster but less thorough than Deep Research.

Best for: AI-assisted video for training, internal comms, sales, and marketing. Vids became free for any Google account holder in April 2026 (with rate-limited Veo 3.1 generation).

• Free Veo 3.1 clip generation for any Google account; full features for Workspace and AI Pro/Ultra
• Custom Lyria 3 music generation, AI avatars / talking heads, multi-track timeline, and template library
• AI scripting, scene generation, and voiceovers from a brief
• Native sharing in Drive, brand-kit support, end-to-end clip generation
• Up to 1,000 Veo videos/month on Google AI Ultra and Workspace AI Ultra

Limitations: Generation quotas vary by plan. Workspace Business plans get full AI features at no extra cost only through at least May 31, 2026. 3-minute audio track cap. Not a substitute for a full NLE (Premiere, DaVinci, Final Cut) on complex edits.

Best for: On-device, edge, fine-tunable, or fully self-hosted deployments. Gemma is Google DeepMind's open-weight family — the open counterpart to closed Gemini.

• Gemma 4 (April 2026) — now released under Apache 2.0, the most permissive license Google has shipped. Positioned as "byte-for-byte the most capable open models"
• Gemma 3 (March 2025) — 1B (text-only), 4B/12B/27B (vision-language); 140+ languages; runs on a single consumer GPU
• Distributed via Hugging Face, Kaggle, Vertex Model Garden, and Ollama
• Self-host for offline workloads, regulated environments, or to avoid per-token costs at scale
• Fine-tunable on your own data without a Google data-handling agreement

Limitations: Smaller context windows than Gemini 3.1 Pro. Less capable on long-horizon reasoning. No built-in tool/agent runtime, no Search grounding — you assemble the inference and orchestration stack. Gemma 3's older terms include Google's prohibited-use policy; only Gemma 4 is true Apache 2.0.