Google Antigravity 2.0 — A Deliberate Lock-In Play

Google isn’t competing with Cursor anymore. At I/O 2026 on May 19, Google shipped Antigravity 2.0 — a standalone desktop app, a Go-based CLI, an SDK for custom agent workflows, and Managed Agents baked into the Gemini API — all running on Gemini 3.5 Flash at 289 tokens per second. Meanwhile, the Gemini CLI that developers adopted just months ago sunsets for non-enterprise users on June 18, 2026. The message is clear: this is not a coding assistant. It is a full developer platform, and Google wants every agent you build to speak Gemini-native.

TL;DR

• What: Google launched Antigravity 2.0 at I/O 2026 — desktop app, Go CLI, SDK, and Managed Agents in the Gemini API, replacing Gemini CLI entirely
• Sunset: Gemini CLI dies June 18 for non-enterprise users — hard migration deadline
• Pricing: New $100/month AI Ultra tier (5x limits), reduced $200 tier (20x limits), $20/month Pro gets full platform access
• Signal: Google is building a vertically integrated agent pipeline that routes through Google Cloud — this is a lock-in play, not a tool launch

Google Antigravity 2.0 — What Happened

Google announced Antigravity 2.0 as a three-surface platform: a standalone desktop application built around agent orchestration, a CLI tool written in Go that fully replaces Gemini CLI, and an SDK for building custom agent workflows. On top of that, Managed Agents in the Gemini API let developers spin up an isolated Linux agent with a single API call — one that reasons, uses tools, and executes code — powered by the same Antigravity harness and available through the Interactions API and Google AI Studio.

The architecture is genuinely different from sequential agent runners. A manager agent breaks tasks into subtasks and dispatches parallel specialized subagents — one writes code, another runs terminal commands, a third tests in the browser. Google’s keynote demo claimed 93 separate subagents processed billions of tokens to build an OS framework in 12 hours for under $1,000 in compute costs. Those are Google-claimed metrics from a keynote stage, not independently audited results, so treat them as a ceiling rather than a guarantee.

The Gemini CLI sunset is the forcing function. On June 18, 2026, Gemini CLI and Gemini Code Assist IDE extensions stop serving requests for Google AI Pro and Ultra subscribers. If your organization uses Gemini CLI via a Gemini Code Assist Standard or Enterprise license, or through Google Cloud’s GitHub integration, access remains unchanged. For everyone else — indie devs, small teams, anyone on the consumer tiers — this is a hard cutover with less than 30 days notice.

Pricing lands at three tiers. Google AI Pro at $20/month gives full access to the Antigravity platform — desktop, CLI, and SDK. A new AI Ultra tier at $100/month provides 5x higher usage limits than Pro. The top tier dropped from $250 to $200/month with 20x limits. Competitive with Claude Code’s Pro pricing on paper, but Gemini API calls beyond quota bill separately through Google Cloud, which is the detail that matters most.

Gemini CLI sunsets for non-enterprise users on June 18, 2026. If you’re using it in CI pipelines or local workflows, you have less than 30 days to migrate to the Antigravity CLI or switch providers entirely.

Why This Matters

This is the third time in two months I’ve written about a major AI vendor shipping the same playbook: give developers a coding assistant, then expand it into a platform, then make the platform the default path for building agents on your cloud. Anthropic did it with Managed Agents. AWS did it with AgentCore. Now Google is doing it with Antigravity 2.0, and they’re doing it faster and more aggressively than either competitor.

The SDK is the key piece to watch. A desktop app and CLI compete with Cursor and Claude Code on the surface where developers write their own code. An SDK competes at a different layer entirely — it’s the interface where your CI pipeline, your deployment automation, and your custom agent tooling all get wired to a specific vendor’s runtime. Once your team builds internal tools on the Antigravity SDK, switching costs compound monthly. Every custom agent, every automated workflow, every test harness that calls the Interactions API becomes a dependency you’d have to rewrite to leave.

The multi-agent architecture makes this stickier than a sequential runner. When your workflow relies on a manager agent dispatching specialized subagents in parallel — where one writes code, another executes terminal commands, a third runs browser tests — you’re not just depending on a model. You’re depending on an orchestration runtime. That’s fundamentally different from calling a completion API where you can swap models behind a proxy like LiteLLM. The orchestration layer is the lock-in layer, and Google just made it a product.

The pricing structure reinforces this. At $20/month for full platform access, the entry point is intentionally low — matching Claude Code Pro’s positioning. But the real revenue comes from two places: the $100–$200/month tiers for heavy usage, and the overflow billing through Google Cloud for API calls beyond quota. This is the classic cloud pricing ratchet. You start with predictable monthly costs, then your usage grows, then your bill starts routing through metered cloud APIs, and suddenly you’re negotiating enterprise agreements with Google Cloud sales.

If you’re evaluating Antigravity 2.0 for a team: audit not just the monthly subscription cost but the projected Gemini API overflow charges. The $20/month Pro tier is the bait — the Google Cloud metered billing is the hook.

The security story is the one genuinely positive development here. Antigravity 2.0 ships with built-in cross-platform terminal sandboxing, credential masking, and hardened Git policies across both the CLI and desktop app. These aren’t bolt-on features — they’re baked into the agent harness itself. Public documentation doesn’t yet allow a direct comparison against Claude Code’s permission modes or Cursor’s isolation model, but the fact that Google is shipping these controls as defaults rather than opt-ins puts pressure on competitors to match. For teams running agents in CI, sandboxing by default is the right call.

The Take

I’ve watched this exact playbook three times now, and the pattern is unmistakable. Ship the IDE to win the developer’s fingers. Ship the SDK to win the developer’s infrastructure. Ship managed agents to win the developer’s cloud bill. Google is running it faster than Anthropic or AWS because they have the distribution advantage — Android Studio, Chrome DevTools, Firebase, Google Cloud all become natural integration points for an agent runtime that already speaks Gemini-native.

The “under $1,000 to build an OS” demo is compelling theater. But the real cost isn’t the demo — it’s 12 months from now, when your CI pipeline calls the Antigravity SDK, your Android builds route through Gemini-optimized tooling, and your custom agents are wired to the Interactions API. At that point, switching to Anthropic or a self-hosted solution isn’t a weekend project. It’s a quarter-long migration.

My recommendation: use Antigravity 2.0 for what it’s good at — fast, parallelized agent tasks where the multi-agent architecture genuinely outperforms sequential runners. But do not build your internal SDK integrations on top of it until the API surface stabilizes and you’ve seen real-world overflow billing data. Keep your agent orchestration behind an abstraction layer. The moment your custom tooling calls Google-specific APIs directly, you’ve made a platform commitment, not a tool choice.

If you’re on Gemini CLI today, migrate before June 18 — but migrate to a neutral harness if you can, not to the replacement Google wants you to adopt by default.