Disclosure: this guide is based on Google’s I/O 2026 announcements, launch documentation and third-party coverage and reviews. We have not run Google Antigravity across a team for months, and figures are attributed in the text.
Google Antigravity started life in November 2025 as an agent-first coding IDE. At I/O on May 19, 2026, Google turned it into something much bigger: a standalone desktop app, a new command-line tool, a developer SDK, a Managed Agents tier inside the Gemini API and an enterprise deployment path. Six months on, it is Google’s single answer to Claude Code, Codex and Cursor.
My honest take: Google Antigravity is the most generous platform in this category for individuals, with a free tier and a Gemini 3.5 Flash default that is cheap and quick. The trade-off is that pricing is spread across Google’s consumer plans rather than spelled out per product, quotas have frustrated some users, and the forced retirement of Gemini CLI left a sour taste. It is worth using. It is worth budgeting carefully.
Table of Contents
- What Google Antigravity Is Now
- The Five Google Antigravity Surfaces
- Google Antigravity Pricing
- Google Antigravity and the End of Gemini CLI
- Managed Agents: The Genuinely New Part
- Gemini 3.5 Flash as the Default
- Google Antigravity vs Claude Code, Codex and Cursor
- The Google Antigravity Catches
- Getting Started Without Wasting Quota
- Rolling Out Google Antigravity Across a Team
- Who Should Use What
- FAQ
- Verdict
- Sources and Further Reading

What Google Antigravity Is Now
The original Google Antigravity was one desktop application: an editor built on the VS Code codebase with agents that could plan, write and test code. Version 2.0 kept that editor but wrapped it in a wider platform. Coverage from Dev.to and Apidog described it the same way, as five distribution shapes for the same agent runtime.
The biggest functional change is parallelism. Version 1.0 ran one agent at a time. Version 2.0 orchestrates multiple agents in parallel, spawning subagents dynamically as a task demands. A single request to refactor a service can split into one agent updating code, another writing tests and a third checking documentation, all at once.
Other additions round it out. Scheduled tasks run in the background using a cron-style schedule command. There is native voice control. And projects can now span repositories, AI Studio apps and Firebase backends, which matters if you build on Google’s stack.
The Five Google Antigravity Surfaces
| Surface | What it is | Who it is for |
|---|---|---|
| Desktop app | Standalone app on macOS, Linux and Windows, editor on the VS Code codebase | Developers who want an agent-first IDE |
| Antigravity CLI (agy) | New terminal tool written in Go | Terminal users, CI pipelines, headless runs |
| SDK | Python, TypeScript and Go libraries | Teams embedding agents in their own products |
| Managed Agents | Hosted agents in the Gemini API, each in an isolated Linux environment | Anyone who wants an agent from one API call |
| Gemini Enterprise Agent Platform | Enterprise deployment on Google Cloud | Organisations with governance requirements |
The surfaces share a runtime, so customisations carry across. Agent Skills, Hooks, Subagents and Extensions from the old Gemini CLI migrated as plugins, which helps teams that had invested in them.
For most individual developers, the choice is between the desktop app and the agy CLI. If you live in an editor, use the app. If you live in a terminal or want agents running in CI, the Go-based CLI is faster than the tool it replaced, according to several reviews.
Google Antigravity Pricing
Google Antigravity does not have its own pricing page. Usage is metered against Google’s AI consumer plans, which is convenient if you already pay for one and confusing if you are trying to compare it with a rival.
| Plan | Price | What you get |
|---|---|---|
| Free | $0 | Individual access with usage limits |
| Google AI Pro | $20/month | Baseline Antigravity usage |
| Google AI Ultra (new) | $100/month | About 5x Pro usage |
| Top tier | $200/month, cut from $250 | Highest usage multiplier |
Tech Insider reported the top-tier cut from $250 to $200 alongside the new $100 plan. Its reading was that Google believes Gemini 3.5 Flash is efficient enough to serve heavy agent workloads profitably at a lower ceiling. That is plausible, and it puts Google’s top plan level with the $200 plans from Anthropic and OpenAI.
Managed Agents and the enterprise tier were not given standalone per-call rates at launch. ToolWorthy noted that cost modelling requires making assumptions about Gemini API billing. If you plan to run agents at scale through the API, ask for a quote rather than extrapolating from the consumer plans.
Google Antigravity and the End of Gemini CLI
This is the part some developers are still annoyed about. Gemini CLI was popular, with more than 100,000 GitHub stars and thousands of merged pull requests. At I/O, Google announced it would be retired in favour of the new Antigravity CLI.
InfoWorld reported the specifics: from June 18, 2026, Gemini CLI and the Gemini Code Assist IDE extensions stopped serving requests for free individual users and for subscribers on Google AI Pro and Ultra. Paid Gemini Code Assist customers kept access. A migration command preserves Agent Skills, Hooks and Subagents.
The migration itself was reasonably smooth by most accounts. The frustration was about the principle. Developers had built workflows on an open, community-backed tool, and a month’s notice to move felt short. It is a useful reminder that free developer tools from large companies can be consolidated whenever strategy changes.
Managed Agents: The Genuinely New Part
Several reviewers made the same point: the desktop app and CLI are polish, and Managed Agents is the structural change. One HTTP request to the Gemini API gives you a reasoning loop, tool use and code execution inside an isolated Linux environment that Google runs for you.
Apidog described the pattern well. You describe a goal, list permitted tools, and Google runs the agent. That turns agentic workflows into something you call the way you call a chat completion, without hosting sandboxes, managing containers or writing the loop yourself.
For product teams, that lowers the cost of trying agent features dramatically. For platform teams, it raises questions about data handling, network access from the sandbox and auditability. At launch, Managed Agents was a preview tier, so treat it as something to prototype on rather than something to put under a production SLA without checking the current terms.
Gemini 3.5 Flash as the Default
Google Antigravity 2.0 switched its default model to Gemini 3.5 Flash. Google advertised it as roughly four times faster than other frontier models and stronger than Gemini 3.1 Pro on coding benchmarks. You can select other models where your plan allows.
A fast default changes how an agent feels. Parallel subagents multiply the number of model calls, so a slow model would make the whole system crawl. Flash keeps multi-agent runs responsive and cheap enough to fit inside consumer plan quotas.
The trade-off is depth. Flash-class models are tuned for speed and cost, and on the hardest reasoning and long-horizon tasks, larger models still lead. For difficult refactors, switching to a stronger model for the planning step is worth the extra quota.
Google Antigravity vs Claude Code, Codex and Cursor
| Tool | Entry price | Default strength | Distinctive feature |
|---|---|---|---|
| Google Antigravity | Free, then $20 | Fast Gemini 3.5 Flash default | Five surfaces, Managed Agents API |
| Claude Code | Included with $20 Claude Pro | Opus and Sonnet 5.5 | Strong agentic coding models |
| Codex | ChatGPT paid plans | GPT-6.1 Sol | Tight ChatGPT Work integration |
| Muse Code | From $5 | Muse Wonder Full on top plans | Lowest headline entry price |
The New Stack noted that Anthropic includes Claude Code with its $20 Pro plan, with Max 5x and Max 20x plans at $100 and $200. Google now matches that ladder almost exactly. The differences are in defaults and ecosystems: Google Antigravity is the natural pick for teams on Firebase and Google Cloud, and its free tier makes it the easiest to try.
On raw model quality for difficult coding, the Claude 5.5 models and GPT-6.1 Sol have stronger published agentic benchmark results than Flash-class models. That does not decide the question, because speed and cost matter for everyday work, but it does suggest using a stronger model for the hardest tasks whichever platform you pick.
The Google Antigravity Catches
Three are worth flagging. First, quotas. One Dev.to review headlined the quota problem, and it is a recurring complaint: parallel agents burn through allowances faster than single-agent tools, and consumer plan limits are not always transparent about what a unit of work costs.
Second, pricing opacity for the API surfaces. Without published per-call rates for Managed Agents at launch, forecasting a production bill is guesswork.
Third, platform risk. The Gemini CLI retirement showed Google will consolidate tools on short timelines. If you build heavily on any one surface, keep your agent logic portable where you can.
Getting Started With Google Antigravity Without Wasting Quota
Because parallel agents spend quota quickly, the first week with Google Antigravity is worth planning rather than improvising.
Begin on the free tier with a small, real repository you know well. Give the agent a single bounded task, such as adding tests to one module, and watch how it splits the work into subagents. That shows you what a typical task costs in your plan’s units before you trust it with anything large.
Write a short project instructions file describing your conventions: how tests are run, which directories are off limits, what the code style is. Agents that start every task by rediscovering the project waste quota on exploration. A clear brief saves that every single time.
Cap parallelism at first. More subagents is not always faster, because they can duplicate work or conflict on the same files. Two or three in parallel is a sensible starting point, raised only when you see tasks that genuinely split cleanly.
If you use the agy CLI in continuous integration, run it on a schedule against a branch rather than on every commit. Scheduled background tasks are one of the new features, and a nightly run that triages failing tests or updates dependencies gives you most of the value at a fraction of the quota.
Finally, review the diff, not the summary. Every agent platform occasionally reports success on work that is incomplete. Treat Google Antigravity’s summaries as a guide to where to look, and let your tests and your own reading decide whether the task is done.
Is an Agent-First Platform Right for Your Work?
Not every developer benefits equally from agent-first tools, and it is worth being honest about that before paying for a higher tier. Agents shine on work that is well specified and easy to verify: adding tests, updating dependencies, applying a known pattern across many files, fixing bugs with a clear reproduction. They struggle with work where the hard part is deciding what to build.
If most of your week is the first kind, an agent platform can save real hours. If most of it is design discussion, debugging subtle production issues or untangling unclear requirements, the gains are smaller, and a good autocomplete may serve you as well as a fleet of parallel agents.
Rolling Out Google Antigravity Across a Team
Individual adoption is easy. Rolling Google Antigravity out to a team of twenty takes a little more thought, mostly about access, cost and consistency.
Decide first which surface is the standard. Mixed use is fine, but support is simpler when most of the team uses either the desktop app or the agy CLI, not both at random. A shared project instructions file checked into each repository keeps agent behaviour consistent regardless of which surface a developer uses.
Then settle the plan question. Because Google Antigravity usage is metered through Google AI consumer plans, a team on individual subscriptions can end up with a mix of Pro, Ultra and free accounts with very different limits. For anything beyond a pilot, the Gemini Enterprise Agent Platform gives central billing and policy control, at the cost of a sales conversation.
Agree on rules for what agents may touch. Production credentials, customer data and deployment pipelines should be out of bounds by default. Google Antigravity can schedule background tasks and run parallel agents, which is exactly why clear limits matter. An agent that can deploy at 3am should only exist because someone decided it should.
Measure before and after. Track cycle time on a sample of tickets, review comments per pull request and the share of agent-written changes that needed rework. Those numbers will tell you whether Google Antigravity is saving time or just moving effort from writing code to reviewing it.
Finally, keep an exit path. Store prompts, skills and hooks in your repositories rather than only inside the tool. If Google consolidates products again, as it did with Gemini CLI, you want your team’s accumulated know-how to move with you.
Who Should Use What
| If you are | Start with | Why | Link |
|---|---|---|---|
| An individual developer trying agents | Google Antigravity free tier | Free access, fast default model | Read the guide |
| A team on Firebase and Google Cloud | Google Antigravity on AI Pro or Ultra | Projects span Google services | See surfaces |
| Building agents into your own product | Managed Agents via the Gemini API | One call, hosted sandbox | See Managed Agents |
| Prioritising the strongest coding model | Claude Code with Sonnet or Opus 5.5 | Higher published agentic scores | See Claude |
Related on this site: for more context, see Claude Code vs Codex vs Cursor, Copilot AI Credits Explained and Gemini 4 Carbon.
Frequently Asked Questions About Google Antigravity
What is Google Antigravity?
Google’s agentic development platform. It launched as an IDE in November 2025 and became a five-part platform at I/O 2026: desktop app, CLI, SDK, Managed Agents and an enterprise tier.
Is Google Antigravity free?
Individuals can use it free with limits. Paid usage comes through Google AI Pro at $20, a new AI Ultra plan at $100 and a top tier cut from $250 to $200 a month.
What happened to Gemini CLI?
Google retired it in favour of the Antigravity CLI. From June 18, 2026, it stopped serving free users and AI Pro and Ultra subscribers; paid Gemini Code Assist customers kept access.
What model does it use?
Gemini 3.5 Flash by default, which Google says is about four times faster than other frontier models. You can choose other models depending on your plan.
What are Managed Agents?
Hosted agents in the Gemini API. One request starts a reasoning loop with tools and code execution in an isolated Linux environment that Google runs.
Is it better than Claude Code?
It is cheaper to start and faster by default. Claude Code pairs with models that have stronger published agentic coding scores. The right choice depends on your stack and how hard your tasks are.
Verdict
Try Google Antigravity if you have not, starting on the free tier. For individuals and teams already on Google’s stack, it is the easiest agent platform to adopt, and Managed Agents is a genuinely useful way to prototype agent features without building infrastructure.
The caveat is cost visibility. Quotas are spread across consumer plans, API surfaces lacked published per-call prices at launch, and the Gemini CLI retirement showed how quickly Google will reshuffle tools. Benchmark claims for the default model are Google’s own.
Next step: run one bounded task on a familiar repository, note how much quota it used, and only then decide whether the $20, $100 or $200 plan fits your real workload.
Sources and Further Reading
- InfoWorld: Google unifies coding tools, for the Gemini CLI retirement details.
- Digital Applied: Antigravity 2.0 deep dive, for the five surfaces and migration notes.
- Tech Insider: pricing changes, for the new tiers.
- Apidog: Antigravity 2, for how Managed Agents works.
- The New Stack: Muse Code pricing, for rival plan comparisons.
On this site: compare tools in AI coding assistants in 2026, connect agents through the best MCP servers, protect them with prompt injection and agent security, and see Google’s video tooling in Google Flow agent prompts.

