Antigravity 2.0: Google takes on Claude Code and Cursor on their own turf

What happened during the keynote

While the mainstream press focused on the Samsung glasses and Gemini Spark, Google quietly launched the most aggressive technical offensive of its keynote: Antigravity 2.0. Varun Mohan (Director of Software Engineering at Google) and Logan Kilpatrick (Google DeepMind) took to the stage to reveal a product that has almost nothing to do with Antigravity 1.0, released in November 2025.

The moment that silenced the room: Mohan asked Antigravity 2.0 to code a complete operating system kernel in parallel, then ran a clone of Doom directly on that freshly created OS. Total cost in tokens consumed: under 1,000 dollars. The audience, made up mostly of developers, probably understood at that moment that the bar had just moved.

As we documented in our full article on Google I/O 2026, this announcement is part of a broader strategy. But Antigravity 2.0 deserves its own treatment, because it directly targets two of the biggest successes in the current AI ecosystem: Anthropic's Claude Code and Cursor.

From standalone editor to complete platform

Antigravity 1.0, launched in November 2025, was a code editor powered by Gemini. Not revolutionary in itself. But above all, as we analysed in our article on the AI startups that outperform, Cursor (built on the same AI-first editor principle) exploded to reach nearly 1 billion in annualised revenue in 2.5 years. Anthropic, for its part, made Claude Code one of its most dynamic sources of enterprise revenue.

Antigravity 2.0 responds to this competition by changing product category entirely. It is no longer an editor. It is a platform with 5 surfaces, all connected to the same "harness" (the underlying agent engine):

Antigravity 2.0 Desktop: a standalone application (macOS, Linux, Windows) that is no longer an IDE but an agent orchestrator. You talk to it in natural language or by voice command, and several agents work in parallel on your request.
Antigravity CLI: a command-line tool that replaces the old Gemini CLI. Recoded in Go for speed. You launch agentic workflows directly from your terminal.
Antigravity SDK: for hosting Antigravity agents on your own infrastructure. Useful for companies that want to keep their code in-house.
Managed Agents in the Gemini API: you call an agentic workflow as simply as you call a chat completion. Google provides the isolated Linux environment.
Gemini Enterprise Agent Platform: the enterprise variant, with governance, audit and compliance.

What is really changing 🎯
Google's strategic insight here matters: the future of AI programming is not a single editor, but a constellation of interconnected tools. The developer will move fluidly between desktop, terminal, mobile and API depending on context. All these surfaces share the same agent engine, so improvements benefit the whole ecosystem at once. That is probably the right read of the market, and it is what has the competition worried.

The engine under the bonnet: Gemini 3.5 Flash everywhere

The entire Antigravity ecosystem runs by default on Gemini 3.5 Flash, the new model we broke down in our Gemini Flash vs GPT-5.5 vs Opus 4.7 comparison. That choice is no accident.

As a reminder, Gemini 3.5 Flash beats its predecessor 3.1 Pro on coding and agentic benchmarks, at 1.50 dollars per million input tokens. Its speed of 289 tokens per second, four times faster than Opus 4.7 and GPT-5.5, becomes critical when running several agents in parallel. Latency compounds: if each agent waits 10 seconds to start, and you launch 5 in parallel, you lose 50 seconds on Claude Opus. On Gemini 3.5 Flash, you lose 12 seconds. On a workflow that runs for several hours, the gain becomes massive.

Gemini 3.5 Flash's score on MCP Atlas (83.6%) and Toolathlon, the multi-tool orchestration benchmarks, is also the highest in the industry. That is exactly the kind of task Antigravity 2.0 runs all day long.

The demo that stuck in people's minds

Let's go back to the highlight of the presentation. Varun Mohan launched an Antigravity 2.0 workflow with a simple request: "Build an OS kernel and run a clone of Doom on it."

Several agents got to work in parallel. While the conference carried on, the screen showed:

One agent working on the kernel's memory architecture
Another handling the input and output drivers
A third coding the file system
A fourth compiling the whole thing
A fifth writing the Doom clone in C

After a few minutes, the system was functional. Mohan launched Doom on this kernel created live, and the game ran. The total cost: under 1,000 dollars in tokens consumed.

Methodological honesty requires a caveat ⚠️
This demo is spectacular, but it is a keynote demo. As we documented in our article on faked AI demos, this kind of presentation may involve cherry-picking or advance preparation. The real question is: what happens when an ordinary developer tries to reproduce it from home? The first user feedback will appear in the coming weeks. Worth watching closely before concluding anything.

The concrete new features of the Desktop App

The centre of gravity of Antigravity 2.0 is the desktop application. Here is what it brings in practice.

Dynamic parallel agents. You launch a complex task, and Antigravity automatically breaks the work down into sub-agents that work in parallel. No manual configuration: the system decides how many sub-agents to launch and how to distribute the work.

Scheduled background tasks. You can schedule agents to work overnight (refactoring code, writing tests, updating dependencies). You find the work done in the morning.

Voice commands. You talk to Antigravity as you would to Spark. "Create me a REST API in Python that handles JWT authentication and an endpoint for user profiles." The agent understands, codes, tests, delivers.

Multi-repository projects. A real difference from Claude Code and Cursor: an Antigravity project can cover several Git repositories simultaneously. If your application has a React frontend, a Python backend and a mobile service in Swift, Antigravity handles all three consistently.

Grouped conversations. Your exchanges with the agents are organised by project, with persistent memory. You pick up where you left off last week.

The new pricing: Google plays the price war

Google has also announced a pricing overhaul that confirms its aggressive intent. Here is the new grid:

Plan	Antigravity (Google)	Claude Code Max (Anthropic)	Codex Pro (OpenAI)
Top tier	$200/month (Ultra)	$200/month	$100/month
Limits compared with the Pro plan	20x Pro	5-20x Pro depending on the period	Variable
Flagship model used	Gemini 3.5 Flash (and 3.5 Pro soon)	Claude Opus 4.7	GPT-5.5

The old Ultra tier was $250 a month. Google has cut it to $200, exactly Anthropic's price for Claude Code Max. The message is clear: "We're going dollar for dollar with the competition, and we're offering 20 times the capacity of our Pro tier."

As a bonus, new Ultra subscribers (and those who already were before 25 May 2026) received $100 in free Antigravity credits. A classic promotion to win over the undecided.

The other strategic integrations

Antigravity 2.0 does not live in isolation. Google has announced several integrations that strengthen the ecosystem lock-in.

Google AI Studio on mobile. A new mobile app (pre-registration opened on 19 May) lets you capture an idea on the move and pick up a working prototype back at the office. You note "I'd like an app that scans barcodes and compares prices", AI Studio starts the work, and you carry on in Antigravity Desktop.

Export to Antigravity from AI Studio. An entire project (with its context, its files, its prompts) can be migrated from AI Studio to Antigravity in one click. No more copy-pasting, no more losing context.

Firebase integration. Google's Firebase backend connects natively to Antigravity. For web and mobile app developers, the "code, test, deploy" workflow becomes continuous.

Native Android from AI Studio. You describe an app, AI Studio generates the Android code, and the Google Play Console integration publishes the app straight to the test track. From idea to published app without leaving the Google ecosystem.

The real threat to Anthropic and Cursor

Let's be clear. Antigravity 2.0 is not going to kill Claude Code or Cursor overnight. But the attack is serious, and it targets precisely their respective weaknesses.

Against Anthropic / Claude Code. Anthropic's weak point is the ecosystem. Claude Code is great for coding, but it has no mobile version, no native cloud integration, no mature enterprise platform. Google is attacking on all three fronts simultaneously. Anthropic will probably have to respond within 6 months, either by building those products itself or by allying more deeply with AWS and Microsoft.

Against Cursor. Cursor is also a great IDE, but it depends on OpenAI and Anthropic APIs for its models. If Google makes Gemini 3.5 Flash 5 times cheaper than Opus 4.7, and the agentic quality is comparable, Cursor has to choose: either shift part of its requests to Gemini (and lose Claude consistency), or lose out on price competitiveness. Neither option is comfortable.

Against Microsoft / GitHub Copilot. Microsoft has Copilot integrated into VS Code, GitHub and Azure. But Copilot is sorely lacking in advanced agentic capabilities in 2026. If Antigravity 2.0 really delivers what was promised, enterprise developers will start to prefer it for complex tasks.

Google's real bet 💰
Antigravity 2.0's goal is probably not to become Google's biggest revenue stream tomorrow. It is to become the default developer infrastructure for the next decade. Once a team codes on Antigravity, deploys on Google Cloud, and uses integrated Firebase and Android Studio, the switching cost becomes prohibitive. It is the same lock-in strategy we described in our article on the OpenAI super-app, but applied to the world of software development.

Our practical verdict

Antigravity 2.0 is probably the most strategic product launch of Google I/O 2026, far more so than the Samsung glasses that made the headlines. Here is what we concretely think of it.

If you are an independent developer: try Antigravity Desktop. It is free to start, it runs on Gemini 3.5 Flash (excellent price-performance ratio), and the desktop version is mature on macOS and Windows. The CLI is still preview quality on Linux, so use it with caution.

If you are an SME or a startup: take a serious look at the Pro plan at 100 dollars a month. You get access to the Desktop, the CLI and the SDK. For a team of 3-5 developers, it is probably the best value for money on the market today.

If you are in a large enterprise: the Gemini Enterprise Agent Platform is worth evaluating, especially if you are already on Google Cloud. Governance, audit and compliance are better thought through than in version 1.0.

If you use Claude Code or Cursor: do not switch everything over straight away. Google's benchmarks are published by Google, and real-world experience may differ. Wait 4-6 weeks to see independent user feedback. But start testing in parallel.

One thing is certain: with Antigravity 2.0, Google is showing that the AI developer tool war is far from over. Anthropic and OpenAI had a two-year head start. Google has just closed half of that gap in a single keynote. The next 18 months will be fascinating to watch.