Everything About Google Gemini 3
Original article: https://mp.weixin.qq.com/s/kH6sFBJ3Zxt94LZ-CGdyPg
Original by Juya 2025-11-19 06:00
Overview
- Google Gemini 3 is Here! #1
- AI Studio Free, but API is Not #2
- Gemini App Gets Major Updates #3
- AI IDE Launches with Free Preview: Google Antigravity #4
- Gemini CLI: Paid API, AI Ultra or Waitlist #5
- Third-Party Platforms All Online #6
Google Gemini 3 is Here! #1
Google has officially released its most intelligent AI model series to date, Gemini 3, featuring the top-performing Gemini 3 Pro and the upcoming Gemini 3 Deep Think mode. It's now available on AI Studio and the Gemini app.
Gemini 3 Pro demonstrates top-tier performance in multiple AI benchmarks, with advanced reasoning, multimodal understanding, agentic coding, and vibe coding capabilities. This release comes with a series of product updates, including integrating Gemini 3 into Google Search (AI Mode) and Gemini app , and launching the new agentic development platform Google Antigravity.
Google CEO Sundar Pichai noted that Gemini 3 is the company's most intelligent model to date, combining breakthroughs from previous generations in native multimodality, long context, and agentic capabilities, while further enhancing reasoning depth and understanding of nuances. Google DeepMind CEO Demis Hassabis stated that the release of Gemini 3 is another important step on the path to AGI.
Gemini 3 Pro performed excellently in multiple AI benchmarks:
- LMArena Leaderboard score reached 1501 Elo , ranking first
- Humanity's Last Exam score of 37.5%
- GPQA Diamond score of 91.9%
- MathArena Apex score of 23.4%
- Multimodal: MMMU-Pro and Video-MMMU scored 81% and 87.6% respectively
- Coding: WebDev Arena ranked first with 1487 Elo
- Terminal-Bench 2.0 score of 54.2%
- SWE-bench Verified score of 76.2%
To further enhance performance, Google also previewed the Gemini 3 Deep Think mode, which achieved higher scores on Humanity's Last Exam and GPQA Diamond in internal testing, and reached 45.1% on ARC-AGI-2. Deep Think mode will first be available to Google AI Ultra subscribers after completing additional safety evaluations.
AI Studio Free, but API is Not #2
Gemini 3 Pro is now available to developers via API, introducing new parameters like thinking level and media resolution, along with detailed pricing, migration guides, and best practices.
Developers can try the model for free in Google AI Studio , but there's currently no free tier for gemini-3-pro-preview in the Gemini API.
API Pricing
| Prompt Length | Input Price (per million tokens) | Output Price (per million tokens) |
|---|---|---|
| ≤ 200k tokens | $2 | $12 |
200k tokens| $4| $18
Thinking Level Parameter
| Thinking Level | Description |
|---|---|
low | Minimizes latency and cost, suitable for simple instruction following, chat, or high-throughput applications. |
medium | Coming soon, not supported at launch. |
high | (Default) Maximizes reasoning depth, first token response time may be significantly longer, but output is more carefully reasoned. |
Media Resolution Parameter
| Media Type | Recommended Resolution | Max Tokens | Use Case |
|---|---|---|---|
| Images | media_resolution_high | 1120 | Ensures highest quality for most image analysis tasks. |
| PDF Documents | media_resolution_medium | 560 | Best for document understanding, quality typically saturates at this level. |
| Regular Videos | media_resolution_low or medium | 70/frame | Sufficient for most action recognition and description tasks. |
| Text-Dense Videos | media_resolution_high | 280/frame | For reading dense text (OCR) or tiny details within video frames. |
Gemini 3 strongly recommends keeping the temperature parameter at its default value of 1.0. Changing this may cause unexpected behavior like loops or performance degradation in complex math or reasoning tasks.
Gemini App Gets Major Updates #3
Google has integrated the new Gemini 3 model into the Gemini app, bringing users smarter reasoning, experimental generative UI, and Gemini Agent functionality for handling multi-step complex tasks.
Gemini 3 uses the most advanced reasoning technology, with more practical, better-formatted, and concise responses. It's the best vibe coding model, with more complete app features built in Canvas. It's also the world's best multimodal understanding model, capable of handling homework photos or lecture note transcriptions.
Gemini 3 Pro is now available globally. Users need to select "Thinking" mode from the model selector. Google AI Plus , Pro , and Ultra subscribers enjoy higher usage limits. Google will provide US college students with a free year of Google AI Pro service, including Gemini 3 access.
New Generative UI Capabilities
- Visual Layout : Generates immersive magazine-style views with photos and interactive modules
- Dynamic Views : Uses Gemini 3's agentic coding capabilities to design and code custom user interfaces in real-time, providing clickable, scrollable interactive experiences
Gemini Agent
Gemini Agent is an experimental feature that handles multi-step complex tasks directly within Gemini. It can connect to Google apps to manage calendars, add reminders, or execute commands like "organize inbox."
This feature is built on Project Mariner insights, powered by Gemini 3's advanced reasoning capabilities, using Deep Research , Canvas , Gmail , Calendar , and other Google Workspace apps, plus real-time web browsing tools. Gemini Agent launches today on the US web version for Google AI Ultra subscribers.
AI IDE Launches with Free Preview: Google Antigravity #4
Google has launched a new agentic development platform, Google Antigravity , which offers an agent-first IDE experience, supporting autonomous planning and execution of complex end-to-end software tasks, and is now available for public preview.
Antigravity 's public preview now supports MacOS , Windows , and Linux. Besides Gemini 3 Pro, it also integrates the Gemini 2.5 Computer Use model for browser control and the Nano Banana image editing model.
Core Features
- Editor Interface : IDE-like experience
- Manager Interface : Agent-first task control center that can generate, orchestrate, and observe multiple agents working in parallel across multiple workspaces
- Supports Gemini 3 Pro, Anthropic Claude Sonnet 4.5, and OpenAI GPT-OSS models
- Unlimited Tab completions and unlimited command requests
- Rate limits refresh every five hours
Gemini CLI: Paid API, AI Ultra or Waitlist #5
Google has integrated the powerful Gemini 3 Pro model into its command-line tool Gemini CLI, bringing developers top-tier reasoning, coding, and tool-calling capabilities, now available to some paid users.
Currently, Google AI Ultra subscribers and users with paid Gemini API keys can use it immediately. Other user groups, including Google AI Pro , Gemini Code Assist Standard , and free tier users, can join the waitlist through a dedicated form for later access.
Activation Steps
| Step | Action |
|---|---|
| 1. Install/Update | Run npm install -g @google/gemini-cli@latest to install the latest version (0.16.x). |
| 2. Enable Preview | Run the /settings command and toggle Preview features to true. |
| 3. Restart and Verify | After restarting CLI, it will default to Gemini 3 Pro. Use /model to manually switch, or /stats to verify current model. |
| 4. Quota Fallback | If daily quota is exhausted, the system will automatically fall back to Gemini 2.5 Pro. |
Third-Party Platforms All Online #6
Gemini 3 Pro has received broad ecosystem support, now integrated into GitHub Copilot , JetBrains IDEs , and multiple other third-party platforms. Its security has also been comprehensively enhanced, making it Google 's safest model to date.
Platforms supporting Gemini 3 Pro include:
- GitHub Copilot (public preview)
- JetBrains IDEs (AI Chat and Junie)
- Cursor
- Manus
- Cline
- Replit
- Kilo code
- flowith
- Genspark
- OpenRouter
On the security front, Gemini 3 is considered Google 's safest model to date, having undergone comprehensive security evaluations showing less sycophancy, stronger resistance to prompt injection, and improved protection against network attacks and other abusive behaviors.