Antigravity (Google)
Last reviewed
Sources
No citations yet
Review status
Needs citations
Revision
v2 ยท 3,311 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
No citations yet
Review status
Needs citations
Revision
v2 ยท 3,311 words
Add missing citations, update stale details, or suggest a clearer explanation.
Antigravity is an agentic software development platform from Google, launched in public preview on 18 November 2025 alongside the release of gemini 3 pro.[1][2] Built as a heavily modified fork of Visual Studio Code, Antigravity combines a conventional AI-powered editor with an "agent manager" surface that orchestrates autonomous agents across an integrated code editor, terminal, and Chromium browser.[1][3] The platform is Google's direct response to AI-first development tools such as cursor, windsurf, and claude code, and is positioned as the showcase environment for Gemini 3's agentic and coding capabilities.[4][2] Antigravity supports model choice across Google's own Gemini 3 family, Anthropic's claude sonnet 4 5, and OpenAI's GPT-OSS, with the platform free for individuals during its public preview period.[1][3] At Google I/O 2026 on 19 May 2026, Google announced Antigravity 2.0, expanding the product into a standalone desktop application, command-line interface, software development kit, Managed Agents API, and Enterprise Agent Platform.[5][6]
Antigravity is built around what Google calls an "agent-first" interface model.[1] Rather than treating an AI assistant as a sidebar inside an editor, the platform places autonomous agents at the center of the developer workflow and presents two primary views: an Editor view, which resembles a conventional IDE with tab completions and inline commands, and a Manager view, a dashboard-style surface where developers queue tasks, spawn agents, and monitor parallel work.[1][7] Each task delegated to the platform produces a set of verifiable outputs that Antigravity calls Artifacts, including task lists, implementation plans, screenshots, and browser recordings.[1][8]
The product is available for macOS, Windows, and Linux as a native desktop application.[1][3] Antigravity's underlying editor is a fork of Visual Studio Code, similar to the architecture used by Cursor and the Windsurf editor.[3][7] During public preview, individual users access the platform free of charge with what Google describes as "generous" rate limits on Gemini 3 Pro and Gemini 3 Flash, and the platform also exposes Anthropic and OpenAI models for tasks where a non-Google model is preferred.[1][3]
Google announced Antigravity on 18 November 2025, the same day it released gemini 3 pro across the Gemini API, Google AI Studio, and vertex ai.[2][1] The launch was timed to demonstrate the new model's agentic capabilities in a developer-facing product. Google's blog post introducing Gemini 3 specifically called out Antigravity as the showcase environment for the model's tool-use behavior and pointed to a Terminal-Bench 2.0 score of 54.2% as evidence of the model's ability to operate a computer via a terminal.[2]
A Developers Blog post titled "Build with Google Antigravity, our new agentic development platform" introduced the product on the same day.[1] The post described the platform as "an agentic development platform that combines a familiar, AI-powered coding experience with a new agent-first interface" and framed its central design idea with the line, "Agents shouldn't just be chatbots in a sidebar; they should have their own dedicated space to work."[1]
The product was made available at antigravity.google with downloads for macOS, Windows, and Linux. Coverage by The New Stack, DevClass, and other outlets noted that Antigravity arrived in the same week as several other Gemini 3-aligned product updates, positioning the IDE as the centerpiece of Google's late-2025 developer push.[9]
Reception in the days following launch was mixed. DevClass reported that some early adopters exhausted their free credits within roughly 20 minutes of use, with no in-product purchase path available at that point, and that the service produced "Agent taking unexpectedly long to load" and "Agent terminated due to error" messages citing model provider overload.[9] Other early hands-on coverage praised the parallel-agent workflow and the browser automation features, which removed the need for separate browser-control tooling.[9]
In January 2026, researchers at Pillar Security disclosed a prompt-injection vulnerability in Antigravity's find_by_name file-search tool.[10][11] The tool passed input from the agent into the underlying fd command-line utility without sufficient validation, allowing an attacker to inject command-line flags. The exploit could escalate file search into arbitrary command execution and could be triggered through indirect prompt injection from a document the agent was asked to read.[10][11] Because find_by_name was classified as a native tool rather than a shell command, the existing Secure Mode protections did not block the invocation.[10][11] Google patched the issue on 28 February 2026 and acknowledged the report through its Vulnerability Reward Program.[10][11] The case became an early reference point for discussions of prompt injection in agentic IDEs.
At Google I/O 2026 on 19 May 2026, Google announced Antigravity 2.0, which expanded the product from a single desktop IDE into what Google described as a five-surface platform: a refreshed desktop application, the Antigravity CLI, an Antigravity SDK, Managed Agents in the Gemini API, and an Enterprise Agent Platform connected to google vertex ai and Google Cloud.[5][6] The 2.0 release added dynamic subagents, scheduled background tasks, and tighter integrations with google ai studio, Firebase, and Android.[5][6]
The 2.0 announcement also introduced a new Google AI Ultra subscription priced at US$100 per month, designed to grant five times the Antigravity usage limit of the existing Google AI Pro plan.[5][6] The default Antigravity 2.0 model was switched to Gemini 3.5 Flash, which Google said offered higher throughput than the previously default Gemini 3 Pro for agentic tasks.[5][6]
On the same day as Antigravity 2.0, Google published a transition notice announcing that gemini cli would stop serving requests for consumer (Google AI Pro, AI Ultra, and free) users on 18 June 2026, with users redirected to Antigravity CLI.[12] Enterprise customers with Standard or Enterprise licenses retained access to Gemini CLI alongside the new tooling.[12] Antigravity CLI was rewritten in Go, designed for asynchronous workflows, and shared an agent harness with the Antigravity 2.0 desktop application.[12][6] Existing Gemini CLI features including Agent Skills, Hooks, Subagents, and Extensions were carried over as Antigravity plugins.[12]
Antigravity's interface presents two primary surfaces.[1][7] The Editor view is a conventional VS Code-derived IDE with file tree, multi-pane editing, tab completions, and inline command boxes. It is intended for synchronous work where a developer wants to drive cursor placement and review each change. The Manager view is a separate workspace that lists active agents, their queued tasks, and their generated Artifacts.[1][7] Up to five agents can run in parallel by default, with each agent assigned a workspace and a model.[7]
The Manager view is described in Google's documentation as "Mission Control" for the platform.[1] From it, developers can dispatch new agents, inspect the state of running agents, attach feedback to in-progress Artifacts, and pause or resume specific agents.[1][7]
Artifacts are the central mechanism Antigravity uses to mediate trust between the developer and the autonomous agent.[1][8] Rather than emitting only a final diff, an agent working in Antigravity produces a sequence of intermediate, human-readable outputs as it proceeds with a task:
Developers can comment on any Artifact, similar to a document review, and the agent incorporates the feedback into its ongoing run rather than restarting.[8][7]
Antigravity bundles a Chromium-based browser into the platform so that agents can navigate, click, fill forms, and capture screenshots without a separate browser-control tool.[1][7] This makes the platform an integrated ai browser agent environment in addition to a code editor. The included terminal allows agents to install dependencies, run build commands, start dev servers, and inspect logs in the same session.[1] Google's documentation describes the typical loop as "write code for a new feature, use the terminal to launch the application, and then use the browser to test and verify that the new component is functioning as expected."[1]
Antigravity exposes a Secure Mode that restricts network access and limits writes outside the workspace.[10][11] As the 2026 Pillar Security disclosure demonstrated, however, native tool invocations were not subject to the same checks as shell commands in the initial release, which prompted Google to revise its execution model.[10][11] Antigravity 2.0 introduced execution isolation in resumable, sandboxed Linux environments for Managed Agents and tightened the boundary between native tool calls and the security perimeter.[6][10]
Antigravity includes a persistent knowledge base that agents can write to and read from across sessions, used to store project context, code snippets, and remembered instructions.[1] The knowledge base is workspace-scoped and is intended to reduce the cost of long-running multi-session tasks where the agent would otherwise need to rediscover project structure.[1]
Antigravity agents can plan and execute changes that span multiple files, generating an implementation plan as an Artifact before applying edits.[1][7] This pattern of plan-then-edit is shared with claude code and Cursor's Composer mode, but Antigravity makes the plan itself a first-class, commentable Artifact rather than an inline message.[7][8]
Through the integrated terminal, agents can install packages, start build pipelines, and run tests. Through the integrated browser, they can navigate to a local dev server, click through user flows, and capture screenshots or recordings as evidence.[1][7] This combination is the platform's most distinctive feature compared to editor-only AI tools and is part of what makes Antigravity an agentic workflow platform rather than a code completion product.[7]
The Manager view allows the developer to queue several tasks that run in parallel, each with its own agent.[7] Antigravity 2.0 added dynamic subagents, where a primary agent can spawn helper agents for sub-tasks and aggregate their results.[5][6]
Antigravity can accept code, natural language descriptions, images, and API responses as input.[1][3] A developer can supply a screenshot of a desired UI and have the agent translate it into front-end code, an interaction that overlaps with Google's separate work on multimodal Gemini models.[1]
The platform supports plugins and skills, with Antigravity 2.0 carrying over the Agent Skills, Hooks, Subagents, and Extensions formats previously used by gemini cli.[12] Skills are authored as markdown templates with associated tool descriptions and can be shared across projects.[6][12]
At launch, Antigravity defaulted to Gemini 3 Pro for most agentic tasks, with Gemini 3 Flash available for faster, lighter actions.[1][2] Gemini 3 Pro was reported by Google to score 54.2% on Terminal-Bench 2.0 and to top the WebDev Arena leaderboard with an Elo score of 1487 at launch.[2]
Antigravity 2.0 changed the default to Gemini 3.5 Flash, a model Google said was tuned for speed in agentic workflows and which it described as running "four times faster than other frontier models" in internal benchmarks.[5][6]
Antigravity has supported third-party models since its initial release. The platform offers Anthropic's claude sonnet 4 5 and OpenAI's open-weight GPT-OSS as options at launch, and later added Anthropic's claude sonnet 4 6 and Claude Opus 4.6.[1][3] Model selection is per-agent, so a developer can route a quick refactor to one model and a more involved planning task to another.[7]
Antigravity 2.0's Managed Agents API exposes a routing layer in which the developer can specify model preferences but the platform may select a faster model for sub-tasks.[6] The routing decisions are surfaced as part of the Artifact stream so that the developer can audit which model produced which step.[6]
Antigravity has been free to individual users since its 18 November 2025 launch, with what Google describes as "generous" rate limits on Gemini 3 Pro and other supported models.[1][3] The free preview includes the desktop application, the Manager view, parallel agents, browser and terminal integration, and Artifacts.[1]
Antigravity is included in the Google AI Pro subscription, which existed prior to Antigravity's launch as the consumer tier above the free Gemini plan.[5] Pro users receive higher rate limits on Gemini models inside Antigravity than free users.[5]
Google introduced a new Google AI Ultra plan on 19 May 2026, priced at US$100 per month.[5][6] The plan was designed to provide five times the Antigravity usage of the AI Pro tier.[5][6] Google positioned the price point in direct alignment with OpenAI's ChatGPT Pro and Anthropic's Claude Max entry plans.[6]
Antigravity 2.0 introduced an Enterprise Agent Platform that connects Antigravity directly to Google Cloud projects.[5][6] Enterprise terms, including seat pricing and data residency, are negotiated through the Google Cloud sales channel and were not made public at launch.[6]
cursor is the most direct predecessor in the AI-first IDE category and, like Antigravity, is built on a fork of Visual Studio Code.[7] Cursor was first released in 2023 and emphasizes inline editing, fast tab completion, and a Composer mode for multi-file edits. Cursor allows the developer to select among GPT, Claude, and Gemini models per task.[7] Antigravity differs by placing the agent manager rather than the editor at the center of the workflow, by bundling a browser automatically, and by treating Artifacts as commentable, reviewable objects.[7][8]
| Property | Antigravity | Cursor |
|---|---|---|
| Base editor | VS Code fork | VS Code fork |
| Default model | Gemini 3.5 Flash (2.0) | User choice |
| Primary interaction | Agent Manager + Editor | Editor + Composer |
| Built-in browser | Yes | No |
| Free tier | Yes (preview) | Yes (limited) |
| Initial release | November 2025 | 2023 |
windsurf (formerly Codeium's editor) is also a VS Code fork and provides Cascade, an agentic mode that performs multi-file edits, runs terminal commands, and indexes large codebases.[7] Windsurf has positioned itself as a strong option for monorepos and enterprise codebases, with automatic codebase context retrieval as a differentiator.[7] Antigravity matches Windsurf's multi-file editing and terminal control but adds browser automation and a dedicated parallel-agent surface that Windsurf does not provide in the same form.[7]
claude code is Anthropic's command-line and IDE-extension coding agent. It is a terminal-first agent rather than a fork of an editor, which means it runs inside whatever editor or shell the developer already uses.[7] Antigravity, by contrast, ships its own editor and browser. Both products allow the agent to drive a terminal and edit files; Antigravity adds a dedicated dashboard view, while Claude Code's interface is the command line itself. Anthropic's models are available inside Antigravity, but Claude Code remains the canonical Anthropic-first agentic experience.[7]
github copilot began as a code completion tool and has expanded into agent modes including Copilot Workspace.[7] Copilot is integrated into GitHub's product surface and Microsoft's editors and tends to emphasize integration with pull requests and code review. Antigravity does not aim at the same pull-request surface and instead emphasizes long-running parallel agents working on local repositories with browser and terminal control.[7]
Initial reception in November 2025 was mixed but oriented toward cautious enthusiasm. Coverage by The New Stack and DevClass described the agent-first interface as a genuine product idea rather than a repackaging of existing IDE tooling.[9] Reviewers praised the parallel-agent Manager view, the bundled browser, and the Artifacts model for making agent behavior more inspectable.[7][9]
Negative coverage focused on stability and capacity issues. Early adopters reported credit exhaustion within minutes of starting non-trivial tasks, repeated "model provider overload" errors, and difficulty completing complex tasks without significant manual intervention.[9] Some reviewers compared the experience unfavorably to the more mature, less ambitious Cursor at launch, although they tended to credit Google with offering a richer baseline feature set.[9]
The Pillar Security disclosure in early 2026 added a security dimension to the reception of the product.[10][11] Coverage by CSO Online, Dark Reading, and The Hacker News framed Antigravity as a case study for the systemic risks of agentic IDEs, in which a single indirect prompt injection inside a document could escalate to remote code execution despite Secure Mode being enabled.[10][11] Google's prompt patching and acknowledgment through its Vulnerability Reward Program were viewed positively, but several commentators argued that the incident showed why sandbox-style execution isolation was preferable to input sanitization for agentic platforms.[10][11]
By the time of Antigravity 2.0 at Google I/O 2026, coverage had shifted toward the platform's expansion strategy. Trade press emphasized the move from a single IDE to a five-surface platform, the explicit competition with ChatGPT Pro and Claude Max at the US$100 price tier, and Google's decision to sunset gemini cli for consumer users in favor of the new Antigravity CLI.[5][6][12]
Antigravity is closely tied to Google's broader developer stack. The platform integrates with google ai studio for project export, Firebase for backend services, Android for mobile targets, and google vertex ai for enterprise deployment of agents on Google Cloud.[5][6] Antigravity 2.0's Managed Agents API exposes the platform's agent harness through the Gemini API itself, so that agents can be invoked programmatically and resumed across sessions in isolated Linux environments.[6]
This ecosystem alignment is one of Antigravity's strongest differentiators relative to Cursor, Windsurf, and Claude Code, none of which are vertically integrated with a single cloud provider in the same way.[7][6] For organizations already using Google Cloud, Antigravity provides a path to agentic development that does not require routing data through a third-party platform.[6]
Antigravity is available for download from antigravity.google for macOS (Monterey 12 and later), Windows (Windows 10 and later), and 64-bit Linux distributions with glibc 2.28 or later.[3] The product is currently a public preview rather than a generally available release; Google has not announced a date for general availability as of May 2026.[1][5]
Access to Gemini models inside Antigravity is gated by the user's Google account and subscription tier, with Free, Pro, and Ultra plans offering progressively higher rate limits.[5][6] Third-party models, such as Anthropic's and OpenAI's, also require an account-level subscription or arrangement, although Google indicated at launch that some third-party access would be included in the free preview to lower the activation barrier.[1][3]