Claude Code vs Gemini CLI, One Got Retired

Claude Code vs Gemini CLI is no longer a live comparison for individual developers. Google stopped Gemini CLI serving Google AI Pro, Ultra and free accounts on June 18, 2026, per its own transition notice, and replaced it with the closed-source Antigravity CLI. Claude Code runs unchanged, included on every paid Claude plan from $20 a month.

Updated September 23, 2026 · by the AITerm team

Look, we had this article half written as a normal comparison. Latency, tool calling, how each one handles a twelve file refactor, the usual. Then we checked whether the free tier numbers were still current, and the answer turned out to be that the tool is not served to individuals anymore.

Honestly, that reframes the whole question. You cannot benchmark your way to a recommendation between a product you can install today and a product whose consumer authentication path returns a permission error.

Is Gemini CLI still available?

For individual developers, no. On June 18, 2026, Gemini CLI and the Gemini Code Assist IDE extensions stopped serving requests for Google AI Pro, Google AI Ultra and free tier individual accounts. Google announced the transition a month earlier, on May 19, framing it as unifying its agent-first development tools around a single platform.

Access did not end for everyone. Three groups kept it: organizations on a Gemini Code Assist Standard or Enterprise license, organizations using Gemini Code Assist through Google Cloud, and anyone authenticating with a paid Gemini or Gemini Enterprise Agent Platform API key.

So the honest answer depends entirely on how you sign in. That distinction is doing a lot of work, and almost nobody writing about this gets it right.

Here's the thing that makes this genuinely confusing rather than merely annoying. The repository was never archived. google-gemini/gemini-cli is still public, still Apache 2.0, sitting on about 107,100 stars and more than 6,400 commits, and its README carries no deprecation banner at all. Google's notice says it stays available to the community with updates for new models, bugs and security fixes for enterprise customers.

You can clone it. You can build it. It will start. It simply will not answer your personal Google account.

Why the comparison you are reading is stale

Search this exact question today and the first page is mostly articles written before May 19, 2026, still weighing a free tier that no longer exists. We checked the top results while researching this piece. The ranking ones compare context windows and pricing tiers without mentioning that one side stopped serving individual accounts three months ago.

That is not laziness on their part, it is how comparison content ages. A post that ranks for a product question gets refreshed when someone notices the numbers moved, and nobody schedules a refresh for a product being switched off.

Two specific claims keep getting repeated, and both deserve a correction.

The first is the context window argument. Every comparison leads with the million token figure as though it settles the matter. In practice, a larger window changes what you can put in context, not what the agent chooses to read.

We have never once lost a refactor because the window was too small. We have lost several because the agent read the wrong four files. Multi-file editing quality tracks retrieval and tool discipline far more than raw window size, and no comparison we found tested that.

The second is the free tier framing. The old Gemini CLI quota, 1,000 model requests a day with up to 60 requests per minute on Flash, was attached to a personal Google account, not to a project or a team.

A maintainer spelled out in the project's own discussions that a single prompt can trigger dozens of model requests, so "1,000 prompts a day" was never what it meant. For a team of five that quota was five separate personal accounts, which is exactly the arrangement nobody wanted to put in writing.

The community threads have the same problem, and they are worth a word because they rank. Search claude code vs gemini cli on Reddit and the highly upvoted comparisons are from before the announcement, arguing about tool calling and loop behaviour on a tool that was still being served at the time. The technical criticism in them is often sharper than the blog posts. The conclusions are answering a question that closed on June 18.

And yes, one more thing people ask constantly: no, Gemini CLI was not based on Claude Code. They are separate implementations from different companies. They look alike because both were built on Ink, the React renderer for terminals, so they redraw and lay out the same way. Shared UI library, not shared lineage.

Claude Code vs Antigravity CLI, the comparison that matters now

The live question is Claude Code against Antigravity CLI, Google's replacement. Antigravity CLI ships as a Go binary called agy, sharing an agent engine with the Antigravity 2.0 desktop application, and Google's notice says Agent Skills, Hooks, Subagents and Extensions carry over, with Extensions becoming Antigravity plugins.

The structural difference is the one worth writing down. Gemini CLI was Apache 2.0 with the source in the open. Its replacement's public repository, google-antigravity/antigravity-cli, distributes installers and releases rather than the source, and sits at roughly 2,400 stars against the 107,100 on the project it replaced.

Installation is a wash, and the symmetry is almost funny. Both are a single curl line, both drop a binary on your PATH, both authenticate on first run.

# Claude Code, macOS, Linux, WSL
curl -fsSL https://claude.ai/install.sh | bash
cd your-project && claude

# or, if you keep your tooling in Homebrew
brew install --cask claude-code

# Antigravity CLI, macOS and Linux
curl -fsSL https://antigravity.google/cli/install.sh | bash
agy

Where they genuinely differ is posture. Claude Code treats the terminal as the primary surface and spreads outward from it, into VS Code, JetBrains, a desktop app, the web and mobile, all reading the same CLAUDE.md and MCP configuration. It also reads an existing AGENTS.md if your repository already has one for other agents, which matters more than it sounds when you are keeping options open.

Antigravity CLI is the terminal outpost of a platform whose centre of gravity is the Antigravity 2.0 desktop application, and it leans on asynchronous background workflows for long tasks. If that platform is where your team already works, the CLI is a reasonable thing to add. If it is not, you are adopting a platform to get a terminal client.

We cover the wider field, including where Codex sits in this, in our comparison of the AI coding agents.

Can you still run the open-source Gemini CLI?

Yes, with a paid API key, and this is the part almost nobody documents properly. Google's notice preserves access for paid Gemini and Gemini Enterprise Agent Platform API keys, so the npm package plus a GEMINI_API_KEY still gives you a working gemini command. What you give up is the monthly subscription model: you now pay per token.

# the consumer sign-in path is the one that stopped working
# a paid API key is not

npm install -g @google/gemini-cli
export GEMINI_API_KEY="your-paid-key"
gemini

# budget on tokens, not on prompts: a maintainer confirmed in the
# project's own discussions that one prompt can fire dozens of
# model requests

Before you wire a team around that, price it honestly. The free tier was 1,000 model requests a day per personal Google account, and moving to a paid key converts a predictable monthly line item into a usage bill that scales with how chatty your agents are. For a team running agents all day, that is not a small difference.

Community forks exist. Several appeared within weeks of the cutoff, and at least one aims to run the same interface against other providers and local models. Turns out that is the reasonable thing to expect from a codebase with 107,100 stars and hundreds of contributors. It is also early: treat any of them as an experiment, not as the thing you put a five person team on this quarter.

If open source in the terminal is the actual requirement, that is a different search, and we wrote it up separately in OpenCode vs Claude Code.

What this costs per month

Claude Code is included on every paid Claude plan. Pro is $20 a month, or $17 billed annually, and Max starts at $100 with a 20x tier above it. Team and Enterprise seats include it too.

Usage runs on rolling five hour session windows with weekly limits on top. Paid subscribers can switch on usage credits to keep working at standard API rates once the included allowance runs out, which is the lever that matters on a deadline.

Google's side no longer has a Gemini CLI line at all. Antigravity CLI rides on Google AI plan tiers at no extra cost, with an individual free tier built around a weekly agent quota rather than the old daily request count. Those tiers were restructured on the same day as the transition announcement, so check Google's current pricing page rather than any figure you read in an article, this one included.

Fair question: is $20 a month worth paying when Google gives you a free tier? For a hobby project on weekends, probably not, and we will not pretend otherwise. For anyone whose agent sessions are load bearing on a weekday, the monthly subscription is not the expensive part of the setup. The expensive part is the morning your tooling stops answering and you have not planned for it.

Which terminal agent should you install?

Start from how you work rather than from a benchmark table, because the benchmark lead changes with every model release and your workflow does not. Four cases cover most of the people who ask us.

The short version: install Claude Code if your work depends on the session finishing, and keep a second agent configured so that a vendor decision is an inconvenience rather than a week.

The lesson we took from watching a CLI disappear

We got this wrong ourselves, so let us go first. AITerm shipped session presets and launch profiles with Gemini CLI wired in, and our onboarding documentation used it as one of the examples of piloting several agents at once. After June 18 we spent nine days pulling it back out: presets, profile defaults, the docs, the screenshots that showed it running in a pane.

Nine days is not a catastrophe. It is also nine days we spent on a vendor's roadmap decision rather than on our own, and the reason it cost that much is that we had treated an open source license as a stability guarantee.

Which brings us to the position we will actually defend. Apache 2.0 protected the code and did nothing for the people using it. The repository is still there, still permissively licensed, still accepting commits. None of that helped anyone whose gemini command started returning a permission error, because what Google switched off was not the code, it was the service the code talks to.

You're probably thinking that is an argument for self-hosting everything, and it is not. The lesson is narrower and more useful: when a tool's value comes from a hosted model behind it, the license tells you who can fork the client, not who can keep working tomorrow. Read the terms of service with the same attention you give the license file.

In practice the portable pieces are boring and they are what saves you. An instruction file in plain markdown that more than one agent reads. MCP server definitions committed to the repository rather than living in someone's home directory. A project that builds with one command, so any agent can verify its own work without being told how.

Two agents, two panes, one screen

Everything above assumes the thing we actually build for: you end up running more than one agent, and the second one is your insurance policy. That is not a hypothetical. In our anonymous daily ping, the median is around three simultaneous agent sessions per user, and that number did not go down when one vendor left the consumer market.

Three sessions is already past the point where tab hunting works. An agent waiting for approval looks exactly like an agent that is thinking, unless something on your screen distinguishes them, and the one that finished twenty minutes ago is the one you notice last.

So that is the friction AITerm is built against. Native macOS panes with a live state per agent session, whichever CLI is running inside them, isolated worktrees so two agents do not fight over the same files, and provider quota windows on screen instead of discovered at the worst moment. Twenty panes streaming agent output sit around 140 MB of resident memory on an M2 Air, with keystroke latency under 8 ms, which is the part a web-stack terminal cannot hold once the output gets dense.

The useful property here is that AITerm does not care which agent won your comparison. It runs claude, codex, agy or anything else that draws to a terminal, which means the next time a vendor retires a binary you change one line in a preset. We wrote up the two-agent workflow in detail in Codex vs Claude Code.

Going further

Sources

1. An important update: Transitioning Gemini CLI to Antigravity CLI, google-gemini/gemini-cli discussion #27274, May 2026, the official mirror on Google's own repository of the Google Developers Blog announcement of May 19, 2026: Gemini CLI and the Gemini Code Assist IDE extensions stop serving Google AI Pro, Ultra and free tier users on June 18, 2026; continued access for Gemini Code Assist Standard and Enterprise licenses, Google Cloud organizations, and paid Gemini and Gemini Enterprise Agent Platform API keys; Antigravity CLI built in Go with asynchronous background workflows; Agent Skills, Hooks, Subagents and Extensions carried over as Antigravity plugins; the Apache 2.0 repository remaining available with updates for new models, bugs and security fixes for enterprise customers; and the stated rationale of unifying agent-first development tools on a single platform.
2. Important Update: Gemini CLI Has Stopped Serving Requests for Individual Accounts, google-gemini/gemini-cli discussion #28017, June 18, 2026: the cutoff taking effect for Google AI Pro, Ultra and free tier individual accounts while enterprise licenses and API key authentication continue, the PERMISSION_DENIED (code 403) eligibility error users hit, the platform-specific Antigravity CLI install commands, the claim that skills, MCP servers, agents and memory files migrate automatically, and community reports of lost conversation history, missing features in the replacement and reduced model quotas.
3. google-gemini/gemini-cli, GitHub repository, consulted September 23, 2026: approximately 107,100 stars, Apache License 2.0, more than 6,400 commits on main, weekly stable releases, and no deprecation or transition notice in the README.
4. Clarification on Daily Request Limit, google-gemini/gemini-cli discussion #4122: the free tier quota of 1,000 model requests per day with up to 60 requests per minute on Flash, and the maintainer clarification that a prompt entry does not equate to one model request because a single prompt can trigger dozens of them.
5. google-antigravity/antigravity-cli, GitHub repository, consulted September 23, 2026: the agy binary, install commands for macOS, Linux and Windows, roughly 2,400 stars, a repository distributing installers and releases rather than full source, a terminal interface sharing an agent engine with Antigravity 2.0, and documented multi-step reasoning, multi-file editing, tool calling, persistent history, session export and remote SSH sessions.
6. Plans and Pricing, Anthropic, consulted September 23, 2026: Claude Code excluded from Free and included on Pro at $20 a month or $17 billed annually, Max from $100 a month with 5x and 20x options, Team and Enterprise seats, rolling five hour session windows with weekly limits, and usage credits at standard API rates once the included allowance is spent.
7. Claude Code overview, Claude Code documentation, 2026: the install commands for macOS, Linux, WSL and Windows, the Homebrew and WinGet paths, the terminal, VS Code, JetBrains, desktop and web surfaces sharing one engine and the same CLAUDE.md, settings and MCP servers, and the note that Claude Code can read an existing AGENTS.md on its own or alongside CLAUDE.md.
8. Our approach to reviewing and merging pull requests, google-gemini/gemini-cli discussion #6900: the project's public contribution process, describing code contributions as the lifeblood of the project, against a reported total of roughly 6,000 merged pull requests from hundreds of community contributors over the project's first year.