Google Antigravity 2.0

Google used I/O 2026 to turn Antigravity from a single editor into a platform. Antigravity 2.0, announced May 19, is now a four-surface system: the Antigravity IDE (a VS Code fork), a standalone desktop app for orchestrating agents, a Go-based CLI, and a Python SDK for building custom agents on Google’s coding stack. The pitch is no longer “a better autocomplete” but “a place to run fleets of agents,” and on that ambition it lands harder than anything else with a free tier.

The desktop app is the centrepiece. It lets you spin up multiple agents and run tasks in parallel, design custom subagent workflows, and schedule jobs that execute in the background without you babysitting them. That last part matters: most agent tools today still assume you are sitting in front of the terminal approving each step. Antigravity 2.0 leans toward fire-and-return, which is the direction the whole category is moving but few have shipped cleanly.

The standout feature is the built-in browser agent. It can open a page, click interactive elements, toggle DevTools, and run a visual QA loop on frontend changes — checking that the thing it just built actually renders and behaves — without you writing a line of Playwright or Cypress. For frontend work this closes a loop that Cursor and Claude Code still hand back to the developer. If you have ever watched an agent confidently “fix” a layout it never looked at, the value here is obvious.

Under the hood the default model is Gemini 3.5 Flash, which scores 76.2% on Terminal-Bench 2.1. That is competitive but not class-leading: Claude Opus 4.8, the model behind Claude Code, sits at 74.6% on the same Terminal-Bench but 88.6% on SWE-bench Verified, and in day-to-day use developers still describe Claude Code as the tool that gets correct code on the first pass more often. Antigravity’s bet is that speed plus orchestration plus a free tier beats raw single-shot accuracy. For exploratory and parallelisable work that bet is reasonable. For one hard bug in a large codebase, a stronger single model still wins.

Pricing is where Antigravity is genuinely disruptive. There is a real free tier with rate limits, Pro at $20/month, and an Ultra tier that Google has been repricing aggressively — down from $250 to $200 for 20x the Pro limits, plus a new $100 tier at 5x Pro limits. Compared with GitHub Copilot’s June 1 move to usage-based credits (1 credit = $0.01, premium requests retired), Antigravity’s flat tiers are easier to reason about, and the free tier is unusually capable for a product that runs background agents.

The launch itself was rough, and PixlRun is not going to paper over it. The 2.0 update was pushed automatically to existing users, and for a significant slice of the developer community it removed the built-in code editor and broke working environments. Rate-limit complaints followed almost immediately. None of this is fatal — it is the kind of thing a fast-moving team fixes in a point release — but it means the reliability reputation is, as of mid-June, entirely unestablished. If you depend on uninterrupted access for client work today, treat Antigravity 2.0 as a second tool, not your only one.

Where it fits: Antigravity 2.0 is for developers who want to orchestrate several agents at once, who do meaningful frontend work that benefits from the browser QA loop, and who are already in or comfortable joining the Gemini ecosystem. The SDK makes it interesting for teams that want to build their own internal agents on Google’s harness rather than rolling everything from scratch. If you want the single most reliable agent for hard backend problems, Claude Code remains the safer pick; if you want maximum model freedom with no vendor lock-in, an open-source agent like Kilo Code is the counter-argument. But for free, multi-agent, frontend-aware development, nothing else on the market currently offers this much surface area.

The honest verdict: this is the most ambitious thing Google has shipped in developer tooling, and the architecture is right for where agentic coding is heading. The execution caught up to the ambition a beat too late. Give it a point release or two, and it could be the default free agent platform. Today, it is a powerful tool you should adopt with a backup plan.

How we would use it. Antigravity 2.0 earns a spot as a secondary agent for parallelisable frontend work: scaffold several components at once, let the browser agent run its visual QA pass, and keep a stronger single-model tool in reserve for the one gnarly backend bug. The free tier makes that low-risk to try. We would not yet trust it as the only tool on a billable deadline, precisely because the auto-update that broke editors is days old in the rear-view mirror and the rate-limit ceiling under heavy use is still being mapped. PixlRun’s read: adopt it now to learn the orchestration model and the SDK, watch the next two point releases for stability, and promote it to primary only once the reliability reputation catches up to the genuinely strong architecture Google has shipped here.