At Google I/O on May 19, Google announced Antigravity 2.0 — a complete rebuild of the agent platform that first shipped in November 2025. The headline numbers are real: 76.2% SWE-bench Verified, isolated Linux sandboxes, a Go-based CLI, a public SDK, and a Managed Agents API that lets you spin up a full reasoning agent with a single call. Google is no longer trailing in the agent platform conversation. But it is not leading either, and the reasons why reveal something important about where this competition is actually going.
What Antigravity 2.0 Actually Ships#
The 2.0 release is architecturally distinct from its predecessor. The original Antigravity was a desktop app — useful for demos, limited for production workflows. The new version is a platform stack:
Standalone desktop application. The rebuilt desktop app acts as an orchestration hub: run multiple agents in parallel, spin up dynamic subagents for split workflows, schedule tasks to execute in the background. The demo at I/O showed parallel agents building a working OS core from scratch for under $1,000 in token costs, then running a live Doom clone on top of it.
Antigravity CLI. Written in Go. This is the most significant architectural acknowledgment Google has made: terminal-native workflows matter. The CLI exposes the same agent capabilities as the desktop app without requiring the GUI. Developers who live in the terminal now have a first-class Google-native option.
Antigravity SDK. A public API for hosting custom agents on third-party infrastructure. This is Google’s answer to Claude Code’s programmatic invocation story — you can now embed Antigravity-powered agents in your own application rather than running Google’s desktop app.
Managed Agents API. A single API call spins up an autonomous agent that reasons, uses tools, and executes code inside an isolated Linux environment. The execution harness runs on Gemini 3.5 Flash; the planning layer uses Gemini 3 Pro.
Pricing restructure. Google introduced an AI Ultra tier at $100/month (5x Pro limits) and dropped the premium AI Ultra plan from $250 to $200/month (20x Pro limits). The pricing change signals that Antigravity usage is expected to scale meaningfully — the old price points were appropriate for occasional use, not daily driver agentic workflows.
The Benchmark Reality#
76.2% SWE-bench Verified is a legitimate result. It is ahead of GPT-5.4 on that benchmark, trails GPT-5.5 “Spud” (82.7% Terminal-Bench 2.0, though SWE-bench Verified comparison is complicated by OpenAI retiring that benchmark), and sits about 11 points below Claude Opus 4.7’s 87.6% SWE-bench Verified.
The more relevant comparison is not the raw score but what drives it. Antigravity’s managed execution layer uses Gemini 3.5 Flash as the default agent harness. Flash is a speed-optimized model — it produces more steps per minute at lower cost than Pro, which matters for multi-agent parallelism. The tradeoff is reasoning depth on genuinely novel problems. That is an architecture choice, not a flaw, but it shapes where Antigravity wins (broad exploratory tasks, parallel worktree workflows) and where it loses (deep algorithmic debugging, novel architecture decisions).
What Google Gets Right#
They acknowledged the terminal. The original Antigravity was IDE-adjacent. Shipping a Go CLI is Google admitting that Claude Code’s terminal-native model is not a niche preference but a workflow requirement for serious engineering. The CLI supports piping, scripting, and batch invocation — the primitives developers actually need to integrate agents into CI/CD pipelines, not just interactive sessions.
The managed execution sandbox is well-designed. Isolated Linux environments that agents can read, write, and execute in without touching the host machine is the correct security architecture. Each agent run is ephemeral; state persistence happens through explicit handoff rather than implicitly. This is meaningfully safer than tools that operate directly on your local filesystem.
The SDK is a real thing. Most agent platforms ship a demo API and call it a developer experience. The Antigravity SDK supports custom agent hosting on third-party infrastructure with documented lifecycle management. That is table-stakes for enterprise adoption.
Where Antigravity Still Falls Short#
It is Gemini-only. The agent runtime runs on Gemini 3.5 Flash and Gemini 3 Pro. There is no multi-model routing, no ability to substitute a different model for specific subtasks, no Claude fallback when Gemini underperforms on a particular type of problem. For teams that have already built workflows around Claude Opus 4.7’s reasoning quality on complex tasks, this is a non-starter.
Claude Code, by contrast, routes through Anthropic models but is infrastructure-agnostic — it runs on AWS Bedrock, Google Cloud Vertex, and direct API, and its Managed Agents API supports custom harnesses that teams can swap out. The composability is not incidental; it is the architecture.
Google Cloud gravitational pull. The Managed Agents API runs on Google’s infrastructure. The Gemini Enterprise Agent Platform, which underpins Antigravity at the enterprise tier, integrates deeply with BigQuery, Cloud Storage, Vertex AI Pipelines, and Google Workspace. That is genuinely valuable if your stack is Google Cloud native. It is friction if it is not. Claude Code’s Routines run on Anthropic’s infrastructure and have no hyperscaler dependency; they run identically whether you are AWS-heavy, Azure-heavy, or multi-cloud.
The MCP ecosystem gap. Claude Code operates in a world of 6,400+ MCP servers and a standard adopted by OpenAI, Microsoft, and the Linux Foundation. MCP servers are how agents connect to everything: Jira, Salesforce, Slack, databases, internal APIs. Antigravity 2.0 ships its own tool connector model. It is not incompatible with MCP — Google has published an MCP-compatible Antigravity connector spec — but the ecosystem depth is an order of magnitude thinner. You are not getting 6,400 community-built connectors on day one.
The Competitive Picture#
Google has made the most serious attempt yet by a hyperscaler to build a genuine terminal-native agent platform. The CLI, the SDK, the managed execution sandbox — these are products built by engineers who understood what they were missing in the original version. The demo at I/O was not marketing; it was engineering.
But Antigravity 2.0 is still solving a local problem: how do you build a great agentic development platform on top of Gemini and Google Cloud? Claude Code is solving a different, harder problem: how do you build the infrastructure for autonomous agentic development that works regardless of which model, which cloud, and which team runs it?
That difference shows up in the details. Multi-model routing. CLAUDE.md organizational memory that travels with projects, not accounts. Agent Teams that spawn up to 15 parallel subagents with isolated worktrees and a mailbox communication model. /ultrareview as a dedicated cloud session with separate context and compute. Routines that run on Anthropic’s infrastructure without your machine being on. These are not features; they are the architecture of a system designed for agents running without human supervision.
Antigravity 2.0 is an excellent interactive agent platform. Claude Code is building toward agents that run while you sleep. Those are not the same product.
Sources:
- I/O 2026 developer highlights: Antigravity, Gemini API, AI Studio — Google Blog
- Google launches Antigravity 2.0 at I/O 2026 — TechCrunch
- Google Launches Antigravity 2.0: Standalone Agent-First Platform with CLI, SDK, Managed Execution — MarkTechPost
- Google Antigravity 2.0: Agentic Dev Platform — Blockchain Council
- Google I/O 2026: Antigravity 2.0 Deep Dive — jangwook.net