Antigravity 2.0: Five Surfaces, One Apology

Google Antigravity 2.0 launched at I/O 2026 this morning as a standalone agent platform — not an IDE fork, not a VS Code plugin, but a first-class desktop application built entirely around orchestrating parallel AI agents. This post anchors to Google I/O 2026, whose keynote landed earlier today (May 19, 2026), and covers every surface that shipped, the pricing reset that directly targets Claude Max, and the honest counter-narrative of what went wrong in the first 72 hours.

The architectural bet is straightforward: agentic work — decomposing large tasks, running parallel subtasks, scheduling background jobs, orchestrating browser and code-execution tools — deserves a home outside the IDE. Antigravity 2.0 is that home. Whether it earns that designation in practice is a question the rocky five-day rollout has not yet settled.

This guide covers what actually shipped across all five surfaces, the dynamic subagent architecture that most coverage misses, the pricing reset decoded, the five-day launch drama, the June 18 Gemini CLI migration deadline, and a full feature matrix versus Claude Code 1.3 and Cursor Composer 2.5. Everything is sourced from the official Google I/O 2026 developer highlights, TechCrunch, PiunikaWeb's post-launch reporting, and the Google Developers Blog transition notice.

Unlike Antigravity 1.0 (a VS Code fork that launched as a single product), 2.0 is a coordinated platform release across five distinct developer surfaces. According to TechCrunch's reporting, all five went live on May 19, 2026. The default model across every surface is Gemini 3.5 Flash, which Google claims outperforms Gemini 3.1 Pro on coding and agentic benchmarks while running at 289 tokens per second output speed — a vendor figure, noted for attribution.

The platform framing matters because it signals how Google intends to compete with Anthropic's Code with Claude platform (which landed the same morning in London) and with OpenAI's Responses API. Antigravity 2.0 is not a model release — it is a harness release. The model is Gemini 3.5 Flash; the product is the surfaces that expose it.

Antigravity 1.0 launched November 18, 2025 as a VS Code fork — our Antigravity 1.0 guide covers that baseline in detail. Manager View could run 5 parallel agents, Chrome integration read DOM and wrote actions, and the default model was Gemini 3 Pro. It was an IDE that could do agentic things.

Antigravity 2.0 inverts that framing: it is an agent orchestration platform that also ships an IDE. The standalone desktop app has two operating modes — Editor and Agent Manager — and the Agent Manager is the primary surface. According to a Google Cloud engineer's writeup on parallel agents, Agent Manager is a mission-control dashboard where developers spawn, monitor, and interact with multiple agents asynchronously across workspaces.

The tension in this pivot is what caused the launch-week backlash. The initial 2.0 release shipped without a prominent IDE path — Google had bet that most Antigravity users wanted the agent interface, not the editor. The developer community disagreed loudly enough that Google added an "Open IDE" button and "Install IDE" prompt in a May 23 patch. The pivot was right in direction; the communication of it was not.

Antigravity 2.0 ships four official slash commands, documented in detail by DataCamp's Antigravity CLI tutorial. Each command maps to a distinct agent execution mode — they are not just shortcuts, they are the primary way a user signals intent to the orchestrator. Understanding which command to use determines whether the agent asks questions, executes autonomously, runs on a schedule, or brings in a browser.

The slash-command surface is where Antigravity 2.0 diverges most visibly from Claude Code's interaction model. Claude Code has a rich command library and plan mode, but its commands are generally scoped to the current task context. Antigravity's /schedule command introduces persistent background automation as a first-class concept — an agent that keeps running after you close the desktop app is a materially different product than an agent that stops when the session ends.

Most coverage of Antigravity 2.0 focuses on the desktop app and pricing. The architectural difference that matters most for developers building on these tools is the subagent model — and almost no publication has called it out explicitly.

In Antigravity 2.0, a main orchestrator agent receives a complex task and dynamically decomposes it into subtasks. It then spawns specialized subagents (examples from demo recordings include frontend/backend/testing splits, and data-cleaner/data-analyzer/ data-visualizer pipelines) and dispatches subtasks to them in parallel. Each subagent runs in an isolated context — it has its own tool access, its own working memory, and no shared state with sibling subagents by default. The orchestrator aggregates results.

This is meaningfully different from how Claude Code 1.3 handles subagents. In Claude Code, subagents are statically defined as markdown files in .claude/agents/*.md — they are authored ahead of time, invoked explicitly, and one level deep: a subagent cannot spawn further subagents. The decomposition logic lives entirely with the human developer who writes the agent definition files. Antigravity's orchestrator performs that decomposition at runtime. For teams familiar with Claude Code 1.3's subagent model, this is not a minor difference in UX — it is a different architecture for task decomposition.

The practical implication: Antigravity is better suited to open-ended tasks where the correct decomposition is not known in advance. Claude Code's static subagent model gives more precise control to developers who know exactly what each agent should do. Neither is universally better — the right choice depends on whether the bottleneck in your workflow is decomposition speed or decomposition precision.

The Antigravity CLI is a full replacement for the Gemini CLI, built from scratch in Go. According to the Google Developers Blog transition notice, the Go rewrite makes it "snappier and more responsive" than the prior Node.js-based Gemini CLI. The CLI invocation shorthand is agy — the full binary is antigravity, but all documentation uses the shorter form.

Install and auth (macOS)

Installation is via Homebrew. The full install, auth, and optional migration sequence:

# Install
brew install google/antigravity/antigravity

# Authenticate
antigravity auth login

# Migrate from Gemini CLI (preserves Skills, Hooks, Subagents)
antigravity migrate --from-gemini-cli

Linux installation is available but described as preview-quality; the macOS and Windows experiences are documented as more stable for production use. Windows install uses Winget or the downloadable installer from the Antigravity website.

Running agents from the CLI

Agent tasks run via antigravity agent run. The CLI accepts a natural-language task string, a --repo flag pointing to the working directory, and an optional --model flag to override the default Gemini 3.5 Flash:

antigravity agent run "refactor the rate-limit middleware" \
  --repo ./services/api \
  --model gemini-3.5-flash

Scheduled tasks use cron syntax directly: agy schedule "daily standup summary" --cron "0 9 * * *". Google frames scheduled background automation as "a first-class place in the editor" — the agent keeps running after the desktop app is closed, and results surface in the Agent Manager view on next open. This is the feature that most distinguishes the Antigravity CLI from both the Gemini CLI it replaces and from competing tools that require external cron infrastructure.

Two pricing moves landed simultaneously at I/O, and reading them together reveals the strategy. The existing AI Ultra+ tier dropped from $250 to $200 per month — a 20% cut on the product that was already live. On the same day, Google launched a new $100/mo AI Ultra entry tier. According to The Next Web's pricing analysis, the $100 tier delivers 5x Pro usage limits and the $200 tier delivers 20x Pro limits.

The competitive read: $100 is the exact price point of Claude Max ($100/mo) and ChatGPT Pro ($100/mo). Google needed a tier at that level to win comparison shopping. The $250 cut to $200 simultaneously signals that the old top tier was not converting at $250 — if it were selling, Google would have added the $100 tier without touching the existing ceiling. Both moves arrived together, which makes the strategic intent unusually legible.

Early AI Ultra subscribers before May 25, 2026 receive $100 in Antigravity credits as a limited-time launch promotion, per the official Google I/O developer highlights. Pay-as-you-go credit packs are also available at $25 per pack ($0.01 per credit), per early third-party reporting. Compute-based billing reportedly replaces daily prompt limits, with quotas refreshing every 5 hours — though that cadence is sourced to a single secondary outlet and should be verified against Google's current documentation.

The Antigravity 2.0 launch was not a clean win. Understanding what went wrong in the first five days is relevant to anyone planning to adopt the platform — it reveals the gap between Google's vision of how developers should use Antigravity and how they actually did.

May 19 — launch day: The initial release shipped with a chat interface as the primary entry point and without a prominent path back to the IDE editor. Developers who had been using Antigravity 1.0 as an IDE found themselves dropped into what PiunikaWeb's Dwayne Cubbins described as a "barebones chat interface" — the IDE hadn't been removed, but it was not surfaced clearly. The developer reaction was fast and negative.

May 22 — global quota reset: Google reset weekly Gemini usage limits across all paid accounts. This was positioned as a goodwill gesture in response to the backlash, though the timing also coincided with the normal weekly quota cycle. The combination of a rocky launch and a quota reset as an implicit apology is unusual for a Google product release.

May 23 — "Open IDE" patch:Google shipped a visible "Open IDE" button and an "Install IDE" prompt in the desktop app. A migration tool to restore settings, extensions, and keybindings from Antigravity 1.0 also shipped in this update. Varun Mohan, Google's Director of Software Engineering for Antigravity, publicly acknowledged the communication failure: Google "should have been clearer from the start."

The launch-week timeline is worth tracking. The IDE/chat-interface tension reflects a genuine product-direction bet: Google believes agent orchestration is the primary workflow, IDE editing is secondary. Most developers on day one disagreed. Whether that gap closes as the user base adapts — or whether Google retreats further toward the IDE — will shape Antigravity's trajectory through the rest of 2026.

The Gemini CLI is not being updated alongside Antigravity 2.0 — it is being retired. Per the Google Developers Blog transition notice, AI Pro and AI Ultra customers lose access to the Gemini CLI on June 18, 2026. The only exception is paid Gemini Code Assist customers, who retain access beyond that date.

The migration is a forced calendar event, not an optional upgrade. If your CI/CD pipelines, developer tooling, or automated workflows depend on the Gemini CLI, you need to migrate before June 18 or those workflows will break.

The practical migration path is straightforward. The Antigravity CLI preserves all the primitives that made Gemini CLI useful: Agent Skills (reusable capability definitions), Hooks (event-triggered actions), Subagent definitions, and Extensions (now called Antigravity plugins in 2.0 terminology). The antigravity migrate --from-gemini-cli command handles the translation automatically. Test your existing workflows in the Antigravity CLI before June 18 — the DataCamp Antigravity CLI tutorial is a good starting point for understanding the new command structure.

Teams using the Managed Agents API are not affected by the CLI sunset — that surface operates via the standard Gemini API and has no dependency on the CLI toolchain.

The Cursor Composer 2.5 launched yesterday (May 18), putting three major agent coding platforms in direct comparison within 48 hours. The matrix below covers the dimensions that matter most for teams choosing a primary agent platform — not marketing claims, but the structural differences in how each tool handles subagents, scheduling, IDE integration, and model support. This is the first post-I/O 2026 comparison putting all three side-by-side with the GUI-vs-TUI distinction explicit.

For a broader landscape view including Codex CLI and Replit, see our roundup of the Claude Code, Cursor, Codex, and Replit ecosystem.

The honest summary: Antigravity 2.0 is the only agent platform from a hyperscaler with a native desktop GUI, the only one with first-class cron-style scheduling in the CLI, and the only one with an orchestrator that decomposes tasks dynamically at runtime. It is also the platform that launched and apologized within 72 hours, and whose IDE integration required a post-launch patch to be discoverable.

Claude Code 1.3 remains the more precise tool for developers who know exactly what their subagents should do and want fine-grained control over decomposition. Cursor Composer 2.5, which landed yesterday, occupies the middle ground — GUI-native, cloud agents with worktree isolation, and a more predictable launch history. For teams evaluating all three, the right frame is not "which is best" but "which architecture matches the task decomposition challenge I actually have." If you need help mapping that to your stack, our team covers exactly this in our AI transformation engagements.

Google did not provide a formal Antigravity 2.1 roadmap at I/O, but the May 23 "Open IDE" patch is the first of what will likely be a series of rapid follow-on updates. The IDE/chat interface tension that defined launch week suggests the near-term roadmap will focus on making the Editor mode more prominent without retreating from the agent-first architectural bet. A migration tool that restores extensions and keybindings is already live; expect additional compatibility and discoverability improvements in the weeks ahead.

The bigger signal on the model side: Gemini 3.5 Pro is framed as a June 2026 arrival in I/O materials. Antigravity 2.0 currently ships on Gemini 3.5 Flash, which Google claims outperforms Gemini 3.1 Pro on coding and agentic tasks while running at higher throughput. Gemini 3.5 Pro — the full-capability frontier model in the 3.5 family — would substantially raise the performance ceiling for Antigravity's most demanding agent workloads. The Gemini 3.5 Flash benchmarks and API guide covers the current model's capabilities in depth.

The competitive context: Code with Claude landed the same morning as Antigravity 2.0, Cursor Composer 2.5 landed yesterday, and OpenAI's Responses API is the third pole in the agent platform race. The next 60 days will determine whether Antigravity's five-surface launch translates into developer adoption or whether the rocky rollout cedes ground to Claude Code's more stable iteration cadence. The Google I/O 2026 hub at our complete I/O announcement guide covers the broader platform context including Gemini Spark, AI Mode, and the full developer surface stack.