Gemini 3 Released - It Changes Everything
Gemini 3 Pro is the best AI model yet released, state-of-the-art across a range of tasks and modalities and raising the bar on AI performance. Google is taking the AI lead, and there is no wall.
Google Releases Gemini 3
We’re beginning the Gemini 3 era by releasing Gemini 3 Pro in preview and making it available today across a suite of Google products so you can use it in your daily life to learn, build and plan anything. - Google
After months of anticipation and weeks of rumors and preview drops, Google released Gemini 3 on November 18th. Google is calling the Gemini release “a new era of intelligence” and they are not understating things. Just in time for the 3rd anniversary of the ChatGPT moment, the Gemini 3 release delivers the latest world’s best frontier AI model, achieving state-of-the-art intelligence and performance across many tasks.
In launching Gemini 3, Google released Gemini 3 Pro in preview and made it available across their product suite:
Gemini 3 is in the Gemini App, which has been upgraded to include a “new type of interface” that adapts to users queries. The Gemini App has also added a Gemini Agent tool for Ultra members, which completes complex, multi-step tasks.
For developers, Gemini 3 was added to Google Studio and Gemini CLI so developers can start building with Gemini 3.
They also introduced Antigravity, an agentic AI coding IDE that they call “a new agent-first development experience.”
Gemini 3 Pro is available via API at a price of $2 per 1M tokens input, $12 per 1M tokens output for under 200k context, increasing to $4 / $18 for input / output on over 200k token contexts. This is an increase from Gemini 2.5 Pro pricing.
Google added Gemini 3 into Google Search, which pro and Ultra users can access Gemini 3 Pro by invoking “thinking” in Google’s AI mode in Search.
Google introduced Gemini 3 Deep Think, augmenting the reasoning in Gemini 3 Pro to push it further.
The Gemini 3 release also continues the ongoing shift from AI as conversational chatbot to AI as agentic assistant that automates your tasks. Google isn’t pitching Gemini 3 as a conversationalist, but as an agentic assistant to help you “learn, build, and plan anything.”
They released Gemini 3 Pro in preview, so we can expect to have incremental updates and Flash and Flash-lite versions in coming weeks and months, just as they did with Gemini 2 and 2.5.
Gemini 3 Pro is SOTA on Benchmarks
It’s the best model in the world for multimodal understanding and our most powerful agentic and vibe coding model yet, delivering richer visualizations and deeper interactivity — all built on a foundation of state-of-the-art reasoning. -Google
Across the board, Gemini 3 Pro achieves state-of-the-art AI benchmark results on reasoning, math, and agentic tasks. It achieves number one on LMArena leaderboards for Text chat, WebDev, and Vision understanding. Most importantly, it shines on reasoning and raw intelligence.
The busy chart of benchmark results below shows that Gemini 3 isn’t just leading on practically all benchmarks, it shows huge leaps in some areas and saturated some difficult benchmarks:
Gemini 3 blows past all other AI models on visual reasoning with 31.1% on ARC-AGI2. It’s SOTA on visual understanding, with 81.0% on MMMU-Pro and 87% on Video-MMMU.
Gemini 3 is strong on math and science. It saturates AIME 2025 with 100% (with tools) and scoring 91.0% on GPQA Diamond. On Humanity’s Last Exam, it gets 45.8% (with search and code execution), well beyond competitors.
On agentic tasks, Gemini 3 gets SOTA scores on Vending-Bench 2 and tau-bench (85.4%).
On coding, Gemini 3 Pro is competitive but not a leader, getting 76.2% on SWE-Bench verified, just behind Claude Sonnet 4.5 and GPT-5.1. However, Gemini 3 got leading scores on Terminal bench 2.0 (54.2%) and LiveCodeBench Pro (2,439).
Gemini 3 Pro is Solid on Vibes
Judging by benchmarks, Gemini 3 is the most profound advance on AI performance in a new AI model release since perhaps GPT-4. However, real-world use determines how it stacks up as an AI model. So, do the vibe tests reveal a solid AI model or any ‘bench-maxxing’?
Most online reviews have been very favorable and confirm Gemini 3’s high performance. AICodeKing found Gemini 3 Pro scored 100% on his evaluations, coding up chessboards and Rust programs, generating SVGs and solving math problems.
Theo called it “the best model ever made” and shared feedback from users on X. Ayush Sharma used over 5M tokens on Gemini 3 Pro and says Google really cooked with Gemini 3 Pro:
Best model for UI Tasks. Tool calling is really great too. Best at one shotting things. Speed is nicer too. 1M context window great for larger projects… Really good at writing also.
His negative points: Gemini 3 Pro did less planning, used less web search, and did not follow instructions as well as GPT5.1. Other reviews noted hallucinations haven’t improved relative to Gemini 2.5 Pro.
Gemini 3 has powerful abilities to generate user interfaces and visual games. Wes Roth shared examples of Gemini 3 coded visual interfaces and games: Minecraft Hippo farm; 3D Lego editor; Pacman on a 3D world; even a Doom-like game in a single shot. Demis Hassabis got in on the video game vibe-coding fun, sharing a Gemini 3 Pro-generated rendition of his own Theme Park video game from the 90s.
Gemini 3 Pro has strengths in visual processing and large context handling that make it uniquely skilled in analyzing and digesting video and user interfaces. It’s the best tool for visual and complex context workflows.
Finally, while Gemini 3 Pro might not beat Claude or GPT-5.1 in persuasive writing or conversational nuance, it holds its own for many writing tasks overall. The 1501 LMArena score attests to its overall capability as an AI model for daily conversational AI use.
Google Antigravity is another AI IDE
We want Antigravity to be the home base for software development in the era of agents. - Google
Google released Antigravity, an agentic AI coding editor that offers a slew of agentic coding features. Antigravity’s “Editor view” offers AI-assisted tab autocompletion, natural language code commands in-line, and a context-aware configurable AI coding agent in the side panel.
If this sounds like a clone of Cursor or Windsurf, that’s because it pretty much is. Antigravity is a fork of VS code like those other AI coding tools and shares most of the features and interfaces. Google aqui-hired the top talent from Windsurf earlier this year, and Antigravity appears to be the fruit of that acquisition.
In addition to Gemini 3 (pro high and pro low) as its primary coding model, it also supports GPT-OSS and Claude Sonnet 4.5 as models. So far, it only supports those specific alternatives and is not yet open to local AI models. Hopefully, they will add ability to use local models soon.
Antigravity goes beyond simply being an AI coding assistant, with autonomous AI coding agents that can plan, design, code, and verify software updates:
We also believe agents deserve a form factor that exposes this autonomy optimally and allows users to interact with them more asynchronously. So, in addition to the IDE-like Editor surface, we are introducing an agent-first Manager surface, which flips the paradigm of agents being embedded within surfaces to one where the surfaces are embedded into the agent.
The Manager interface makes the user interaction behave more like Devin or Manus agent systems. Antigravity allows agents to propose terminal commands, code diffs, and browser actions that engineers approve or reject, transforming AI-driven coding from a keystroke-heavy task to a specification and review process.
While Google opened this up to anyone and offers a free tier, some online reviews complain that rate limits are strict. For software developers, it may be worth a test drive as a Cursor alternative.
Gemini 3 Changes Everything
Google’s release of Gemini 3 marks a new chapter in the AI era.
Gemini 3 Pro is the world’s best AI model overall, extremely useful across a broad range of tasks. AI users should try this on various use-cases and see where it works best. With strength on agentic AI, math, and reasoning, it could be excellent on work-related queries.
Software developers should definitely try Gemini 3 Pro as an alternative AI model, especially in coding interfaces and visual artifacts, where it appears to be more advanced than competitors.
With this release and the release Google has become the leading AI lab, with the broadest and best AI technology suite overall. First-mover OpenAI has advantages in mind-share and market-share, and there is stiff competition from Anthropic and the Chinese labs, but no company has been more productive than Google this year. Google has managed to move faster and is releasing at a relentless pace.
As of publication, Google is already on to their next release: Nano Banana Pro. Google is not slowing down.
There is STILL no wall. The leaps in performance are significant enough to confirm that AI performance scaling has not hit a wall yet. Demis Hassabis says Gemini 3 is “dead on track” with the progress he expected:
It’s (Gemini 3) on right on track of what I’ve been expecting and the trajectory we’ve been on since the beginning of Gemini … We are going to continue that trajectory and we expect that to continue.
Just as Gemini 3 is significantly better than last year’s Gemini 2, we can expect further advances in coming months and years. While Gemini 3 Pro is the best AI model for now, both competitors and Google itself will be beating this AI model soon.
Finally, AI is getting fully agentic. Gemini 3 shows strength in agent benchmarks and agentic use-cases, emphasizing capabilities such as tool-calling and integrating with tools such as embedded AI agent in the Gemini App. Anti-gravity shows the evolution of AI coding from AI assistant to AI agentic coding tool.
Adding an agent to the Gemini app interface completes this evolution to agentic AI. AI apps are not just for chatting anymore.