Google Antigravity 2.0: What Changed in the Agentic IDE

Google Antigravity has become one of the most searched terms by developers in recent hours — and for good reason. At Google I/O 2026, the company transformed what started as an agentic IDE into a full-fledged AI-powered development platform. In this guide, I explain what Google Antigravity 2.0 is, what changed from the first version, and why it affects those who write code every day.

TL;DR — what you need to know

• Google Antigravity is no longer just an IDE: Antigravity 2.0, announced at Google I/O 2026 on 19 May, has become a platform with five usage surfaces.
• The surfaces are the desktop app, Antigravity CLI, SDK, Managed Agents via the Gemini API, and the Enterprise Agent Platform.
• The default model is now Gemini 3.5 Flash, which — according to Google — outperforms Gemini 3.1 Pro on nearly all benchmarks while running four times faster.
• Gemini CLI will be discontinued on 18 June 2026; users must migrate to Antigravity CLI before then.
• There is a free tier, included in Google AI Pro, and an AI Ultra plan at US$100 per month with five times the usage limits.

What is Google Antigravity 2.0

Google Antigravity was born on 18 November 2025, launched alongside the Gemini 3 model. At the time, Google presented it as its first agentic development platform — a term worth unpacking. Agentic describes software that not only suggests code snippets but plans and executes entire tasks on its own: opens files, runs terminal commands, tests in the browser, and checks its own results.

The first version was, in practice, a modified fork of Visual Studio Code, available for Windows, macOS, and Linux. What set Google Antigravity apart from a regular editor was the core idea: agents gained their own surface, the Agent Manager, with direct access to the editor, terminal, and an embedded browser. Instead of typing every line, the agent received a task, created a plan, and returned verifiable artefacts — implementation plans, task lists, walkthroughs, and screenshots showing what it had done.

Antigravity 2.0, announced at Google I/O 2026, takes the next step. It is no longer a simple IDE but becomes a standalone desktop application built entirely around agent orchestration. The IDE still exists within it, but has become just one piece of a larger ecosystem. It is this shift in scope that explains the spike in searches: many people want to know if they need to switch tools.

From IDE to ecosystem: what changed in version 2.0

If version 1.0 was an editor with built-in agents, version 2.0 is a command centre for agents. The difference is not cosmetic. Here is what is new in Google Antigravity:

• Native voice commands. You create and trigger workflows by speaking, without typing — the same feature Google brought to Gmail and Docs.
• Dynamic subagents. A main agent splits a complex task into a tree of specialised subagents that run in parallel. Instead of a queue, you have multiple simultaneous workers.
• Scheduled tasks. You can schedule agents to run in the background — a nightly test battery or a scan for outdated dependencies.
• Ecosystem integrations. Antigravity 2.0 natively integrates with Google AI Studio, Android, and Firebase.

The Agent Manager has also matured. It allows you to observe and coordinate multiple agents asynchronously, tracking each one's progress without blocking your own work. In practice, the developer shifts from writing every line to reviewing and coordinating. It is a habit change as profound as switching programming languages.

This asynchronous model is what most confuses those coming from a traditional editor. You do not wait idly for the agent to finish before doing something else. You launch three or four tasks, monitor them via the Agent Manager, and review each one as it completes. Dynamic subagents take this further: instead of manually breaking down a large problem, the main agent decides how to split the work and how many subagents to create. It works well when the task has independent parts; it stalls when everything depends on everything else.

The five surfaces of Google Antigravity 2.0

The biggest structural change is that Antigravity has become a platform with five usage forms. Each caters to a different developer profile:

Surface	Purpose	Best for
Desktop app	Orchestrate agents with full graphical interface	Daily individual use
Antigravity CLI	Create and run agents via terminal, no GUI	Automation, scripts, and CI/CD
Antigravity SDK	Program agent behaviours and host them anywhere	Teams wanting custom agents
Managed Agents (Gemini API)	Trigger agents with a single API call	Products embedding agents
Enterprise Agent Platform	Connect agents to Google Cloud projects	Companies with corporate governance

The SDK is perhaps the most strategic. It exposes the same agent engine that Google uses internally in its own products, optimised for Gemini models. Managed Agents run inside an isolated Linux environment and accept agent definitions written in markdown files — making behaviour auditable and versionable like any other code.

Another detail worth highlighting about Managed Agents: each agent maintains state between sessions. Files, installed dependencies, and conversation context persist from one day to the next. You can pause a long task today and resume it tomorrow without the agent losing the thread. For something like migrating an entire project, this persistence makes a real difference.

The Enterprise Agent Platform rounds out the offering for corporate customers. It connects agents to Google Cloud projects, allowing you to apply access policies, auditing, and centralised billing — requirements that any medium-sized company demands before unleashing generative AI on engineering teams. For solo developers or small teams, the desktop app and CLI handle almost everything.

Antigravity CLI: the mandatory migration from Gemini CLI

Here is the part with a deadline. Alongside Google Antigravity 2.0, Google announced that the Gemini CLI and Gemini Code Assist IDE extensions will be discontinued. The key dates:

1. 19 May 2026 — Antigravity CLI becomes available to all users.
2. 18 June 2026 — Gemini CLI and Gemini Code Assist extensions stop serving requests on Google AI Pro, Ultra, and free plans.

Those with corporate Standard or Enterprise licences for Gemini Code Assist are not affected for now. For all individual plans, however, migration is not optional — it is just a matter of doing it before or after the old tool stops responding.

The good news is that Antigravity CLI preserves what mattered in Gemini CLI. Agent Skills, Hooks, Subagents, and Extensions — the latter rebranded as Antigravity plugins — continue to work. The common pitfall is leaving everything until 17 June: if you use Gemini CLI in CI pipelines, test the migration early, because any incompatibility will break the build in production at the worst possible moment.

In practice, migration has three steps: install Antigravity CLI, run the import command that brings your Gemini CLI settings, and review the extensions that became plugins. The official guide published by Google covers each case, including integration with GitHub, which will stop accepting new installations after 18 June.

Gemini 3.5 Flash: the models behind Google Antigravity

Google Antigravity 2.0 ships with a new default model: Gemini 3.5 Flash. According to Google, it outperforms Gemini 3.1 Pro — the most advanced model of the previous generation — on nearly all benchmarks while running four times faster. For an agentic workflow, speed is not a luxury: an agent that makes dozens of sequential calls feels every millisecond saved.

But Antigravity is not tied to Gemini. The platform also runs Gemini 3 Pro, Anthropic's Claude Sonnet 4.5, and open models from OpenAI's GPT-OSS line. There is an important technical detail: Gemini models receive exclusive optimisations — prompt caching and tool call optimisation — that third-party models do not benefit from. In other words, you can use Claude or GPT, but Gemini tends to be faster and cheaper within Antigravity.

Why this affects your budget

In agentic workflows, token consumption skyrockets. Every generated plan, every file reread, and every correction attempt costs. Choosing the right model for each task — Flash for volume, Pro for hard reasoning — is no longer a detail but a budget decision. Those who ignore this often discover the cost only when the bill arrives at the end of the month.

How much does Google Antigravity cost: plans and pricing

Google Antigravity 2.0 maintains an accessible entry level, but all prices are charged in US dollars. What is confirmed so far:

Plan	Price	Usage limits
Pro	Included in Google AI Pro	Entry-level limits
AI Ultra	US$100/month	5× the Pro plan limits
Higher Ultra tier	US$200/month	20× the Pro plan limits

The AI Ultra plan at US$100 is the most talked-about new addition — it is the sweet spot for those who use agents intensively without justifying the more expensive tier. This higher tier, at US$200 per month, was actually reduced from a previous price of US$250. As a launch promotion, Google offered US$100 in Antigravity credits for accounts activated by 25 May 2026.

Converting roughly for the Brazilian developer, AI Ultra falls in the range of R$550 to R$600 per month, depending on exchange rates and taxes on international services. This is not a trivial amount — so it is worth measuring consumption on the free tier before upgrading.

Is Google Antigravity worth it? When to use and when to wait

Not every Google I/O novelty needs to enter your workflow tomorrow. Antigravity 2.0 shines in some scenarios and still warrants caution in others.

Worth it when:

• You need to prototype quickly and final quality can be refined later.
• The task is repetitive and well-defined: mass refactoring, dependency updates, writing tests.
• You can parallelise — multiple subagents handling independent parts of the code.
• You are exploring an unfamiliar codebase and want a map before diving in.

Think twice when:

• The CLI is still in preview quality on Linux; in production, validate thoroughly before relying on it.
• The task requires subtle architectural decisions — the agent executes well but does not replace engineering judgement.
• You do not have time to review the output. Unverified agentic code is technical debt disguised as productivity.

An expectation adjustment is in order: part of the technical community sees Google Antigravity less as a competitor capable of retiring Cursor and more as Google's bet to keep developers within the Gemini ecosystem. The verification artefacts — plans, walkthroughs, and screenshots — exist precisely for you to check the agent's work. Use them. Ignoring them throws away the tool's biggest advantage.

In my experience, the most expensive mistake is not the tool generating bad code — it is the developer accepting good code without understanding why it works. When something breaks three months later, no one on the team can explain the original decision. Treat the agent like a very fast junior colleague: great for speed, dangerous without careful review. This mindset applies to Antigravity and any competitor.

What Google Antigravity changes for developers in Brazil

My take, after more than fifteen years building software, is that agentic tools like Google Antigravity change the mechanical part of the work, not the hard part. They accelerate the kind of custom development we deliver daily — from custom Moodle apps to system integrations — but they do not decide what should be built, nor why.

An agent can scaffold a screen in minutes. It does not tell you, however, whether the end user actually needs that screen. Decisions like choosing between the official Moodle app and a custom app remain human judgement, informed by business context that no AI model has.

If you code, the advice is straightforward: install Antigravity 2.0, spend a few hours on the free tier, and form your own opinion before 18 June — the date Gemini CLI goes away. And if your company is evaluating how generative AI fits into the product roadmap, this is a good time to talk to those who are already hands-on. The technology has levelled up; what hasn't changed is the need for someone with judgement at the helm.