Google Antigravity: Models, Pricing, and the agy CLI Explained
🚀 Google Antigravity: A Complete Breakdown
Antigravity, the agent-first IDE Google unveiled alongside Gemini 3 in November 2025, is a multi-model coding environment where you pick between Claude, GPT, and Gemini inside a single tool. This write-up walks through five threads in order: ① the real story on supported models and per-plan usage, ② what users are actually saying, ③ where Google is steering the product, ④ how it compares to and pairs with Claude Code, and ⑤ how to drive it straight from the CLI (agy). Where sources contradict each other, both positions are kept side by side rather than forced into a single verdict.
📅 As of June 2026 · Preview-stage information; figures are subject to change
1. What Antigravity Actually Is
The design of Antigravity was led by the former Windsurf team, whom Google brought on in July 2025 through a $2.4 billion licensing deal. The product forks VSCode and layers an autonomous agent-orchestration layer on top of it. Antigravity 2.0, announced at Google I/O in May 2026, shipped a desktop app together with the first official CLI, agy — cementing its role as the official successor to the existing Gemini CLI.
Its core identity is not autocomplete but parallel agent orchestration. Several agents work at once — one on the API, one on tests, another on the frontend — splitting the work between them, and each leaves behind an Artifact capturing its plan, test results, screenshots, and recordings. The mental model is not "the human approves each line as it goes," but "the agents finish their work and the human reviews after the fact." This matters because the review burden shifts to the end of the loop rather than gating every step, which is what lets multiple agents run unattended in parallel.
flowchart TD
A([User task instruction]) --> B[Agent A
API implementation]
A --> C[Agent B
Test authoring]
A --> D[Agent C
UI generation]
B --> E[Artifact
plan / results / video]
C --> E
D --> E
E --> F([Human reviews after the fact])
style A fill:#3498db,stroke:#2980b9,color:#ffffff
style B fill:#e8f8f5,stroke:#16a085
style C fill:#e8f8f5,stroke:#16a085
style D fill:#e8f8f5,stroke:#16a085
style E fill:#fef9e7,stroke:#f39c12
style F fill:#3498db,stroke:#2980b9,color:#ffffff
🔗 Diagram in brief: A single instruction fans out to multiple agents owning the API, tests, and UI simultaneously; each leaves its output as an Artifact, and the human only reviews at the very end. The point is parallel work plus post-hoc review, not sequential approval.
2. Models and Pricing — Transparency Is the Weakest Link
Supported models (8, as of CLI 1.0.2)
| Provider | Model | CLI label |
|---|---|---|
| Google DeepMind | Gemini 3.5 Flash (Low/Medium/High) | "Gemini 3.5 Flash (Low)", etc. |
| Google DeepMind | Gemini 3.1 Pro (Low/High) | "Gemini 3.1 Pro (Low)", etc. |
| Anthropic | Claude Sonnet 4.6 (Thinking) | "Claude Sonnet 4.6 (Thinking)" |
| Anthropic | Claude Opus 4.6 (Thinking) | "Claude Opus 4.6 (Thinking)" |
| OpenAI | GPT-OSS 120B | "GPT-OSS 120B" |
As you'd expect, Claude and GPT models really are in the picker, and the Models pane shows per-model consumption. The catch is that burn rate varies enormously by model. In one user-reported case, a single Claude Opus 4.6 session ate 635 of 1,000 credits in one shot — the heavier the model, the faster it chews through your allowance. The practical takeaway: treat the frontier "Thinking" models as something you reach for deliberately, not as a default you leave running.
Subscription plans and monthly pricing
| Plan | Price/mo | Limit multiplier | Notes |
|---|---|---|---|
| Free | $0 | Baseline (undisclosed) | Reset cadence disputed |
| AI Pro | $20 | 4× Free | Unclear in real-world terms |
| AI Ultra (Entry) | $100 | 5× Pro | Added May 2026 |
| AI Ultra (Premium) | $200 | 20× Pro | Cut from the previous $250 |
| Add-on credits | $25 | +2,500 credits | Buy when depleted |
Stripped down to the monthly fee alone, the spectrum from free preview to $200 is easy to take in at a glance.
⚠️ The confirmed gap — exact per-plan numbers are "not disclosed"
🔴 This is the crux. Google publishes only the multipliers — "Pro is 4× Free" — while keeping the baseline value itself and the credit-to-token conversion rate undisclosed. A second pass opened the official docs page (antigravity.google/docs/models) directly and got back an empty page, and community write-ups go no further than "model access depends on your plan." The net result: "how many hours or requests of Claude Opus do you get on Pro?" is a gap that cannot be answered from any currently public material.
On top of that, since May 2026 billing is no longer keyed to a fixed number of prompts but to compute consumption — a blend of request complexity × number of active tools × cumulative conversation length. There's no way to estimate cost up front, which widens the uncertainty further. In practice, the only way to measure real consumption is through runtime commands.
/credits # Remaining credits + purchase link /usage # Per-model quota and rate-limit status
⚔️ Conflicting sources (reset cadence): Early official guidance said the limit "resets roughly every 5 hours," but user reports describe what amounts to a "7–10 day lockout." One account also cites weekly available tokens collapsing from 300M → 9M (about a 33× drop) between January and March. This is a point where the official story and lived experience flatly contradict each other.
Visualizing the reported drop in weekly available tokens makes the magnitude clearer.
3. Community Verdict — Promise Acknowledged, Quota and Stability Disappoint
🟢 What people praise
✓ Running five projects at once via the Agent Manager is a workflow rivals don't offer
✓ The built-in Chromium browser lets agents click and scroll through the UI to test it, recording video as they go
✓ Reports that Gemini 3 nailed a complex game clone on the first try — one that other top models had failed
🔴 What people criticize
▶ Quota collapse is the hottest complaint — forums are blanketed with grievances that "renews every 5 hours" turned into an effective 7–10 day lockout
▶ Not recommended for deadline-bound projects — frozen buttons, vanishing sidebar icons, battery drain, no keyboard support in the file explorer, and other UI bugs
▶ No Git Worktrees support, and incompatibility with some VSCode extensions (Svelte, among others)
▶ Most seriously, one reported codebase-deletion incident
In short, the consensus is "the potential is clear, but preview-stage instability and opaque billing are holding it back." It's appealing to early adopters who want to experiment with new capabilities fast, but the risk factors are still too many to make it the primary tool on a production project with a delivery date riding on it.
4. Where Google Is Headed and Where It Shines
The picture Google is painting isn't a drop-in replacement for the Gemini CLI but an orchestration platform where agents talk to one another to finish software. The features added in 2.0 point in that direction.
🔹 Background task scheduling — work overnight with no human in the loop
🔹 Custom subagent workflows — you design the division of labor between agents yourself
🔹 Antigravity SDK — outside developers can build and integrate custom agents
🔹 Google Cloud integration — Firebase, Android Studio, and AI Studio in one pipeline
🔹 The fact that Gemini 3.5 Flash itself was developed with Antigravity reads as a vote of confidence from Google in its own tool
Where it genuinely shines
| Sweet spot | Why |
|---|---|
| Frontend / full-stack prototyping | Browser automation lets the agent handle UI verification too |
| Fast MVP builds | Parallel agents generate API, tests, and UI at once |
| Apps with image assets | Integrated automatic generation of game/UI assets |
| Nonlinear, complex projects | Manager View coordinates many agents in parallel |
Conversely, the prevailing real-world consensus is that backend logic, legacy refactoring, and CI/CD automation are better suited to Claude Code. Put simply, Antigravity is strong at "quickly stamping out visible artifacts" and still weak at the "invisible stability and structure" work.
5. Claude Code (Terminal) vs. Antigravity
| Dimension | ⌨️ Claude Code | 🚀 Antigravity |
|---|---|---|
| Runtime environment | Terminal (alongside your existing editor) | Standalone desktop app + CLI |
| Agent model | Sequential, approve before changes | Parallel multi-agent, tunable autonomy |
| Context window | 200K (1M in beta) | 1M by default |
| Browser integration | None (add via MCP) | Built-in Chromium, video capture |
| Stability | Mature | Preview, many bugs |
| Pricing | From $20/mo (no free tier) | Free preview + $20–$200 |
| Default model | Claude Sonnet 4.6 | Gemini 3.5 Flash |
Claude Code's strengths: you keep your terminal-centric workflow and bolt AI onto it, and it stays stable across complex backend, infrastructure, legacy refactoring, and CI/CD work. Approval-gated sequential editing suits developers who value control over every change.
Antigravity's strengths: it can close the loop on frontend prototyping — including automated browser verification — in a single cycle, and parallel agents compress the wall-clock time on large jobs. Being a free preview also lowers the cost of experimentation.
Using both together in practice — "split by role"
graph LR
A[Antigravity
planning / frontend] --> C[MCP server
GitHub integration]
B[Claude Code
backend / data] --> C
C --> D([Single repo, shipped])
style A fill:#e8f8f5,stroke:#16a085
style B fill:#eafaf1,stroke:#27ae60,color:#1e8449
style C fill:#fef9e7,stroke:#f39c12
style D fill:#3498db,stroke:#2980b9,color:#ffffff
🔗 Diagram in brief: A hybrid setup where Antigravity builds the frontend and Claude Code builds the backend in parallel, then an MCP GitHub server merges both into the same repo. The key is splitting roles so each tool sits in its strength zone.
A concrete scenario: while Antigravity builds the UI for a SaaS dashboard, you implement the billing logic in parallel from a Claude Code terminal, and an MCP GitHub server merges the two outputs into the same repo.
6. Driving Antigravity Models Straight from the CLI
Per primary material from Google Codelabs, agy is verified to be a standalone TUI you can install and run independently of the desktop app. It shares the same agent runtime as the desktop app, so sessions and settings stay in sync.
Install (from the Codelabs source)
# macOS / Linux curl -fsSL https://antigravity.google/cli/install.sh | bash export PATH="$HOME/.local/bin:$PATH" agy --version # Windows PowerShell irm https://antigravity.google/cli/install.ps1 | iex
Binary path: ~/.local/bin/agy (Unix) / %LOCALAPPDATA%\Antigravity\bin\agy.exe (Windows)
Basic usage
agy # Interactive TUI agy -p "prompt" # One-shot query
Switching models — selecting Claude Sonnet/Opus
⚔️ Conflicting sources (CLI syntax): the primary source (R1) gives a short flag with a lowercase model name, like agy -m claude-sonnet-4.6, whereas the second-pass verification (R2: Codelabs, DEV.to) reports agy --model "Claude Sonnet 4.6 (Thinking)" plus an in-session /model followed by arrow-key selection. Because the labels and flags differ, the safe move in practice is to run agy models (v1.0.5+) first to see exactly what labels your current build recognizes.
# R2-verified form (quoted full name) agy --model "Claude Opus 4.6 (Thinking)" agy models # List supported models (v1.0.5+) # In-session switching /model # Pick a model with the arrow keys, then Enter /credits # Remaining credits /usage # Per-model usage
Migrating from the Gemini CLI
agy plugin import gemini # Auto-migrate existing MCP servers, allowed commands, keybindings
The existing Gemini CLI's Hooks, Subagents, Skills, and Extensions have all been ported to agy, and the Gemini CLI is scheduled to be discontinued for AI Pro/Ultra users on June 18, 2026. If you're already on the Gemini CLI, a single migration command carries your environment over intact, so the switching cost is on the low side.
7. Takeaways and Implications
🧠 Model choice: being able to pick Claude Sonnet/Opus 4.6, GPT-OSS 120B, and Gemini 3.x from one tool is the core draw. But heavy models (Opus and the like) burn credits very fast.
💰 Per-plan usage: only multipliers like "Pro = 4× Free" are published; the absolute numbers are unverifiable because the official docs page is empty. The reset cadence also conflicts — official "5 hours" vs. real-world "7–10 days" — so for any work where cost predictability matters, this is a risk right now.
🎯 Sweet spots: strong at frontend/full-stack prototyping, automated browser verification, and fast MVPs. Backend, legacy, and CI/CD favor Claude Code.
⚙️ Recommended setup: the most realistic configuration is a hybrid — keep Claude Code as your primary and bring in Antigravity as a sidekick only for UI prototyping and browser-automation tasks. Being a free preview keeps the experimentation cost low.
🔍 Still to confirm: exact per-plan credit/token limits can only be filled in once the official docs are fleshed out, or measured directly at runtime with /usage and /credits.
Contradictions across research rounds
▶ CLI model-selection syntax: R1 gives agy -m claude-sonnet-4.6, R2 gives agy --model "Claude Sonnet 4.6 (Thinking)" plus arrow-key selection — flags and labels don't match
▶ Quota reset cadence: initial guidance "~every 5 hours" vs. user reports of "an effective 7–10 day lockout"
▶ Per-plan usage figures: the official docs (antigravity.google/docs/models) return an empty page, so they're undisclosed — a confirmed gap that further research could not fill
📚 References
• Google Antigravity official pricing page (antigravity.google/pricing)
• TechCrunch — Antigravity 2.0 announcement (I/O 2026)
• Antigravity Pro quota forum (discuss.ai.google.dev)
• devclass — user backlash over the price increase
• Google Codelabs — Antigravity CLI hands-on
• DEV.to — Antigravity CLI guide / candid review
• DataCamp / Augment Code — Claude Code vs. Antigravity comparison
⚠️ The prices, quotas, and CLI syntax in this article come from preview-stage material and may change over time; some items are unverified because the official docs are empty, and figures conflict across sources. In actual use, confirm the latest details directly with runtime commands (/usage, agy models) and the official pages.
I gather material from a software-development angle, organize it myself, and give it one more check before publishing.
This article is based on publicly available data and sources. Last updated: 2026-06-08
댓글
댓글 쓰기