Google Antigravity Review: Features, Pricing, and Verdict

This Google Antigravity review examines the current 2.0 platform without treating an ambitious agent demo as proof of dependable software delivery. Google has expanded Antigravity from an agent-first IDE into a standalone command center where agents can work with files, run commands, search the web, use Chrome, call external tools, create plans, and delegate to subagents.

The short version: Antigravity 2.0 is a serious option for developers who want multi-agent work, visual artifacts, browser testing, and access to several models in one Google-backed product. Its shifting quota system, fast-moving feature set, and broad local permissions make a measured pilot more sensible than an immediate team-wide switch.

Google Antigravity Review: Quick Verdict

Google Antigravity is most compelling as an orchestration surface. A developer can define a Project, choose local or Git worktree execution, review an implementation plan, and monitor parallel work. The same agent system is available through the desktop application and CLI.

The product is no longer simply a VS Code-style editor with a chatbot. Antigravity 2.0 is a standalone desktop application, although Google still provides an IDE component during installation. That distinction matters: the product's direction is toward supervising agents and reviewing artifacts rather than spending the entire day editing code line by line.

This is an evidence-based review, not a hands-on benchmark. Google's documentation supports the product details; Thoughtworks and current security reporting add independent context. None proves that Antigravity will complete a particular repository task correctly.

What Is Google Antigravity 2.0?

Antigravity 2.0 is Google's desktop platform for launching, monitoring, and coordinating AI agents across coding and broader knowledge-work tasks. An agent can read and write project files, execute system commands, perform web research, interact with Chrome, connect to Model Context Protocol servers, use skills, create artifacts, and invoke subagents.

Projects define the folders and repositories an agent can access. One Project can include several folders, such as separate frontend and backend repositories, and each Project has its own settings and permissions. Conversations can run directly in local folders or in new Git worktrees. Worktrees are the safer choice for parallel coding because each conversation receives an isolated checkout rather than editing the same active files.

The current model documentation lists Gemini 3.6 Flash, Gemini 3.5 Flash, Gemini 3.1 Pro, Claude Sonnet 4.6, Claude Opus 4.6, and GPT-OSS-120B. Google's models and plans pages do not describe third-party access identically, so confirm the live selector before paying for a particular model.

Google Antigravity Features That Matter

Projects, Worktrees, and Parallel Agents

Project-scoped settings let a trusted internal repository use different rules from unfamiliar code. Worktree mode reduces collisions when several agents edit a Git repository at once.

Antigravity 2.0 supports built-in research, browser, and self-style subagents as well as custom agents defined in Markdown. Subagents start with isolated conversation context and can run concurrently. Their tool and security scopes are inherited from the parent, which limits accidental privilege expansion but also means an overly broad parent configuration affects every child.

Google also documents a preview teamwork mode for the top Ultra tier. Parallelism can shorten independent work, but it can multiply token use, duplicate effort, and enlarge the review surface.

Plans, Diffs, and Browser Artifacts

Artifacts are structured outputs such as implementation plans, code diffs, architecture diagrams, images, and browser recordings. Users can leave inline feedback on a plan before the agent changes files and review a visual diff afterward. This creates a better control point than relying on a long chat transcript.

The browser subagent uses a separate Chrome profile. It supports on-demand invocation, Chrome DevTools MCP, and WebM recording. That helps reproduce UI bugs, but one recording does not prove every state, viewport, or flow works.

CLI, Hooks, Skills, and MCP

The Antigravity CLI exposes the same agent-oriented workflow from a terminal. JSON hooks can run local scripts around tool calls, model responses, or loop-stopping events. Skills package reusable instructions, while MCP servers connect external tools and data.

This extensibility also adds supply-chain and permission questions. Review every skill, hook, plugin, and MCP server as operational configuration, especially when it can reach credentials or production data.

Google Antigravity Pricing and Quotas

Google currently provides a baseline quota across individual plans. All tiers receive Gemini model usage, unlimited tab completions, scheduled tasks, and CLI access. Free users receive a weekly baseline; Pro gets a higher quota with five-hour refreshes until a weekly cap; Ultra gets the highest limits.

Google's May 2026 announcement priced Pro at $20 per month, introduced a $100 Ultra tier, and reduced the top Ultra tier to $200. It described their shared Gemini quotas as one, five, and twenty times the Pro allowance, respectively.

Those ratios are more useful than a fixed prompt estimate. Google says quota draw depends on the task, model pricing, token mix, and capacity. Simple tasks may allow more prompts than long agent loops, and limits can change.

Pro and Ultra users can spend purchased AI credits after baseline quota is exhausted. Keep automatic overages off during evaluation, then decide from accepted-task cost. Bring-your-own API keys cannot currently extend the included limits.

Google Antigravity Security and Privacy

Antigravity's default Project configuration bounds files to selected folders and asks before terminal commands. Deny, Ask, and Allow rules control web access, browser interaction, commands, MCP tools, and non-workspace files, with Deny taking precedence.

The important caveat is that read and write operations inside an active Project are auto-allowed by default. A malicious instruction inside a repository can therefore influence edits even when the agent cannot immediately run every command. The Full Machine and Unrestricted presets broaden the potential blast radius further and should not be normal settings for untrusted code.

Terminal sandboxing adds operating-system containment and network restrictions, but Google's documentation describes it as preview functionality on macOS and Linux, with Windows support still developing. The CLI documentation also says its sandbox toggle is disabled by default. Enable containment explicitly, deny network access unless the task needs it, and keep secrets outside agent-readable folders.

The July 2026 GhostApproval research showed why an approval prompt is not automatically trustworthy. Researchers demonstrated that several coding agents, including Antigravity, could show a harmless symlink path while a write reached a sensitive target. Reporting says Google fixed the issue, and no exploitation in the wild was reported. Update promptly, inspect real paths and diffs, and isolate unfamiliar repositories.

For a broader control framework, see Syntax Dispatch's AI agent security guide.

Google Antigravity vs Cursor, Claude Code, and Codex

Antigravity's clearest advantage is its integrated command-center design: Projects, artifacts, browser recordings, worktrees, subagents, desktop and CLI surfaces, and Google models.

Cursor favors editor-centered inline iteration. Claude Code suits terminal-first users, while Codex emphasizes parallel tasks, isolated worktrees, and local and cloud execution. Syntax Dispatch's Claude Code vs Codex comparison examines those workflows.

Google Jules is narrower and more asynchronous, focusing on cloud repository tasks rather than a general local command center.

No comparison should start with a generic leaderboard. Give each tool the same repository setup, bug fix, small feature, and test task. Measure accepted changes, regressions, reviewer minutes, latency, quota use, unsafe actions, and recovery from feedback. Syntax Dispatch's AI coding agents guide offers a broader shortlist.

Who Should Use Google Antigravity?

Antigravity is worth testing for developers and small teams that want multi-agent orchestration, local repositories, browser-assisted verification, and several models in one interface.

It is a weaker fit for buyers who need predictable usage, bring-your-own endpoints, or simple autocomplete. Non-technical users may find its permission, quota, worktree, and sandbox choices heavy.

Start with one disposable or low-risk repository. Use the Default security preset, a new worktree, explicit terminal approval, and the sandbox with restricted network access. Add models, tools, folders, persistent permissions, and paid overages only after the first workflow produces repeatable value.

Conclusion

This Google Antigravity review finds a capable, forward-looking agent platform whose best ideas are Projects, worktree isolation, reviewable artifacts, browser evidence, subagents, and a shared desktop-and-CLI workflow. Antigravity 2.0 makes supervising AI workers feel more deliberate than adding another chat panel to an editor.

The tradeoffs are just as material. Quotas are workload-dependent, pricing and model access continue to evolve, agent extensions expand the trust boundary, and local execution can affect real files before a human understands the full plan. Antigravity deserves a structured pilot, not blind trust.

For teams willing to measure outcomes and configure permissions carefully, Google Antigravity can be a useful orchestration layer. For teams that primarily need inline coding help or predictable usage, a simpler editor or terminal agent may remain the better choice.