Gemini 3.0 and Antigravity
Google (MacRumors, The Verge, Hacker News, model card, Hacker News):
And now we’re introducing Gemini 3, our most intelligent model, that combines all of Gemini’s capabilities together so you can bring any idea to life.
It’s state-of-the-art in reasoning, built to grasp depth and nuance — whether it’s perceiving the subtle clues in a creative idea, or peeling apart the overlapping layers of a difficult problem. Gemini 3 is also much better at figuring out the context and intent behind your request, so you get what you need with less prompting. It’s amazing to think that in just two years, AI has evolved from simply reading text and images to reading the room.
And starting today, we’re shipping Gemini at the scale of Google. That includes Gemini 3 in AI Mode in Search with more complex reasoning and new dynamic experiences. This is the first time we are shipping Gemini in Search on day one. Gemini 3 is also coming today to the Gemini app, to developers in AI Studio and Vertex AI, and in our new agentic development platform, Google Antigravity — more below.
Today, we are introducing Google Antigravity, our new agentic development platform. While the core is a familiar AI-powered IDE experience with the best of Google’s models, Antigravity is evolving the IDE towards an agent-first future with browser control capabilities, asynchronous interaction patterns, and an agent-first product form factor that together, enable agents to autonomously plan and execute complex, end-to-end software tasks.
[…]
We believe Antigravity’s product form factor represents the next fundamental step function in agent-assisted development. Thus, our goal is to channel it into the best product offering possible for end users. In today’s public preview:
- Google Antigravity for individuals at no charge
- Compatibility with MacOS, Linux, and Windows
- Access to Google’s Gemini 3, Anthropic’s Claude Sonnet 4.5 models, and OpenAI’s GPT-OSS within the agent, offering developers model optionality
It’s an Electron app based on Visual Studio Code and requires signing in with a Google account.
A key innovation is how the tool reports its work. It shares task lists, plans, screenshots, and browser recordings to make it easier for the developer to confirm it’s doing the right thing.
In addition to the traditional IDE view it also has a Manager view is designed for controlling multiple coding agents at once.
Too bad they’ll probably kill it in a year 😁
Previously:
- White Label Gemini on Private Cloud Compute
- Nano Banana
- Claude Code Experience
- Google I/O 2025
- Google Is Winning on Every AI Front
- Gemini 2.0
- Google’s AI Search and “Web” View
- Google Vertex AI Search
- The Era of Visual Studio Code
Update (2025-11-20): Google (Hacker News):
Today, we’re introducing Nano Banana Pro (Gemini 3 Pro Image), our new state-of-the art image generation and editing model. Built on Gemini 3 Pro, Nano Banana Pro uses Gemini’s state-of-the-art reasoning and real-world knowledge to visualize information better than ever before.
Update (2025-11-26): PromptArmor (via Hacker News):
An indirect prompt injection in an implementation blog can manipulate Antigravity to invoke a malicious browser subagent in order to steal credentials and sensitive code from a user’s IDE.
Update (2025-12-02): Deep-Hyena492 (Hacker News):
Google Antigravity just deleted the contents of my whole drive.
5 Comments RSS · Twitter · Mastodon
I've been playing around with it for two hours this morning, and it doesn't seem substantially better than other vibecoding tools. It's fast, which is good, but it's still ultimately an incredibly frustrating experience. It very confidently spits out large amounts of code that looks credible but doesn't work, followed by lengthy attempts to fix things while breaking others.
Personally, I also feel like this approach to writing code takes the joy out of programming.
@Plume nailed it. Had the same sad experience when trying to write code with the help of Copilot. Incredibly frustrating.
"An indirect prompt injection in an implementation blog can manipulate Antigravity to invoke a malicious browser subagent in order to steal credentials and sensitive code from a user’s IDE."
Same as every other LLM product that has access to both trusted and untrusted data and the ability to interact with things, including literally every single LLM-focused browser.
"Google Antigravity just deleted the contents of my whole drive."
I was actually wondering about this. By default, Antigravity either sets or encourages you to set (can't remember which it was) Terminal Command Auto Execution to "Auto". "Auto" apparently means that the LLM decides whether the command is dangerous and, if so, asks whether it can execute it. If it thinks it's not dangerous, it just runs it.
The obvious problem here is that *the LLM* decides whether to ask. Since the reason it needs to ask in the first place is that it can't reliably write correct commands, it also can't reliably determine if a command is dangerous. So it was only a matter of time until something like this happened.
Instead, Terminal Command Auto Execution should be set to "Off", and users should add a whitelist of commands that they know the LLM can safely execute.
> Google Antigravity just deleted the contents of my whole drive.
We went from 1) Unix permissions ought to be enough for anyone to 2) no, actually, even within a user account, sandboxing should further constrain what processes can do to 3) it’s an LLM; what could possibly go wrong giving it full disk access in no time.
