TL;DR
Antigravity CLI, Claude Code, and Codex CLI are the three terminal coding agents worth comparing in 2026, and the verdict has held since they were first compared: Claude Code is the pick when code quality comes first. Codex CLI is the value play: a bit leaner on tokens, and the only one of the three that’s open source. Antigravity CLI ties its Gemini models to Google’s opaque “compute effort” quotas, which still make it hard to recommend below the $200/mo Ultra tier. All of this traces back to Google killing Gemini CLI for individual users on June 18, 2026. The open-source tool that hit 100K GitHub stars stopped serving requests unless you hold an enterprise license. Antigravity CLI is its closed-source, Go-based replacement, and launch-era reports had free-tier users locked out for up to seven days after 20-30 minutes of use. Google has since moved to 5-hour refreshes under a weekly cap, with paid AI-credit top-ups. That makes the ceiling explicit, but you still can’t see how close you are to it in tokens.
Pricing, default models and plan limits last verified: November 4, 2026.
Google’s Forced Migration
Google announced Antigravity CLI at I/O 2026 on May 19. The same day, the Gemini CLI deprecation notice went live. The timeline was tight, and it played out exactly as scheduled:
- May 19: Antigravity CLI launched, available to everyone
- June 18: Gemini CLI stopped serving requests for Google AI Pro, Ultra, and free-tier users
- June 18+: Gemini Code Assist for GitHub stopped accepting new installations
Enterprise customers with Gemini Code Assist Standard or Enterprise licenses keep their Gemini CLI access indefinitely. Everyone else had to migrate or switch tools before the cutoff.
The backlash was immediate, and for good reason. Gemini CLI was open source under Apache 2.0. Antigravity CLI isn’t. The GitHub repository contains a changelog, a readme, and a demo GIF, but no actual source code. The rate limits on the replacement are drastically more restrictive than what Gemini CLI users had. And the 30-day migration window was enough time to install a binary, but nowhere near enough to rebuild CI pipelines, port custom extensions, or audit a closed-source tool’s data handling.
Gemini CLI, covered in the May tutorial, worked well for what it was: a Node.js-based terminal agent with Gemini models, MCP support, and a sensible free tier. Antigravity CLI is supposed to be the upgrade. In practice, the experience is more complicated.
Antigravity CLI: What You Actually Get
Installation is straightforward:
# macOS / Linux
curl -fsSL https://antigravity.google/cli/install.sh | bash
# Verify
agy --version
The binary is agy, not antigravity. Authentication uses browser-based Google Sign-In. It opens your default browser, you log in, and the token stores in your system keyring. On SSH sessions, it prints a URL you complete in a local browser.
Antigravity CLI shares its backend with the Antigravity 2.0 desktop app. Conversations started in the CLI can be exported to the desktop GUI. Skills, hooks, MCP servers, and context files (GEMINI.md, AGENTS.md) all carry over from Gemini CLI, though the config paths change. They have also moved once since launch; per Agentpedia’s deep dive (updated October 30, 2026, against Google’s current MCP docs), the layout is now:
# MCP config (global)
~/.gemini/config/mcp_config.json
# MCP config (workspace)
.agents/mcp_config.json
# Skills (global / workspace)
~/.gemini/antigravity-cli/skills/
.agents/skills/
# User settings
~/.gemini/antigravity-cli/settings.json
# Migrate existing Gemini CLI extensions
agy plugin import gemini
There’s also a headless mode now: agy -p "prompt" (aliases --print and --prompt) runs a single prompt non-interactively, and --output-format switches between text, json and stream-json. The CLI changelog shows Google hardening it through the 1.x releases: a structured exit code 3 with an AGY_ERROR line on stderr (1.2.6), lower CPU use (1.4.7), and in 1.5.0 -p waits for background tasks unless you pass --print-timeout. For CI, authenticate with a GEMINI_API_KEY rather than the browser sign-in.
Agentpedia describes the Go binary as faster to start and lighter on memory than the Node.js-based Gemini CLI, though it publishes no timings, and no independent benchmark of startup time has surfaced. The TUI is keyboard-driven. On pure ergonomics, it’s an improvement.
Where things fall apart is the rate limiting.
The Rate Limit Problem
Antigravity CLI uses a metric Google calls “compute effort” instead of transparent token counts. A simple question might consume 1 unit. A multi-file refactoring task might eat 50. The UI shows a capacity percentage, but at launch it was unreliable: users on Google’s developer forums, as XDA reported, saw the meter read “100% capacity remaining” right before a 7-day lockout triggered.
The tiers as of November 2026:
| Tier | Price | Models | Quota mechanics |
|---|---|---|---|
| Free | $0 | Gemini 3.8 Flash | Weekly starter quota; launch-era reports of 7-day lockouts after 20-30 min of use |
| AI Pro | $19.99/mo | 3.8 Flash + Gemini 4 Argon | 5-hour refresh under a weekly limit; Argon has its own weekly quota; 6-10 day cooldowns reported at launch |
| AI Ultra (entry) | $100/mo | 3.8 Flash + Argon | 5x Pro limits, weekly limit still applies |
| AI Ultra (top) | $200/mo | 3.8 Flash + Argon | 20x Pro limits, weekly limit still applies |
The quota has been reworked since this comparison first ran. In July 2026, Google moved Antigravity from daily prompt caps to “compute used”, which refreshes every 5 hours until a weekly limit is reached, and let Pro and Ultra subscribers buy pay-as-you-go AI credits once they hit it. Gemini 4 Argon reached Antigravity on October 13, first for Ultra and then for paying Pro subscribers, with a separate weekly allowance whose size Google hasn’t published. Credits soften the ceiling. They don’t fix the unit: usage is still metered in compute effort, not tokens, so you can’t predict when a heavy session will hit the wall.
Weekly caps aren’t unique to Google, though. Anthropic’s usage-limits page says Claude Pro and Max ($100/mo, 5x Pro) combine a 5-hour session limit with a weekly limit. Ultra ($200/mo, which was called Max 20x until late October) gives 10x Pro usage per week and drops the 5-hour limit for verified accounts. OpenAI’s Codex pricing page lists Pro plans from $100 and says they currently have no five-hour limit. The difference is visibility: Claude Code and Codex meter usage you can see, and both price per token on the API. Antigravity’s compute-effort meter is a black box.
The free-tier wall is the harshest. Launch-era reports had 20-30 minutes of active use locking a free account out for up to seven days. The quota is tuned for casual one-off questions, and it breaks down during real development sessions, where a single multi-file refactor burns through the allowance.
Head-to-Head Comparison
Architecture, Models, and Benchmarks
| Feature | Antigravity CLI | Claude Code | Codex CLI |
|---|---|---|---|
| Language | Go | TypeScript (Bun) | Rust |
| Models (Nov 2026) | Gemini 3.8 Flash (all plans); Gemini 4 Argon (paid) | Opus 5.5 default on all paid plans and the API | GPT-6 family; docs’ config example uses GPT-6.1 Sol |
| Open source | No | No | Yes (Apache 2.0) |
| SWE-bench Verified (spring models) | 80.6% (Gemini 3.1 Pro) | ~81% (Opus 4.6) | ~80% (GPT-5.4) |
| Terminal-Bench 2.0 (spring models) | 68.5% (Gemini 3.1 Pro) | 65.4% (Opus 4.6) | 77.3% (GPT-5.3-Codex) |
| Token efficiency | Unknown (opaque) | ~277K tokens/workflow (Composio, Aug 2026) | ~224K tokens/workflow |
| Context files | GEMINI.md, AGENTS.md | CLAUDE.md | AGENTS.md |
| MCP support | Yes | Yes | Yes |
| Subagents | Yes (pro / flash subagents) | Yes | Yes (parallel) |
| Headless / CI mode | Yes (agy -p) | Yes | Yes |
The benchmark rows are vendor-reported model scores from the spring 2026 generation, the models these tools ran when this comparison first appeared in May. They measure the model, not the CLI harness around it. Every vendor has shipped at least one newer generation since. The current cross-vendor head-to-head is Terminal-Bench 4.0, where Anthropic’s Opus 5.5 launch post reports Opus 5.5 at 66.4% and quotes OpenAI’s 57.9% for GPT-6 Astra. No comparable Google figure for Gemini 4 Argon has been published by Google itself.
The open-source gap is significant. Codex CLI’s full codebase lives on GitHub under Apache 2.0. Claude Code isn’t open source either, but Anthropic publishes transparent token pricing, so you can predict costs. Antigravity CLI is closed-source with opaque “compute effort” billing. You can’t inspect what the tool sends to Google’s servers, and you can’t predict what a task will cost before you run it.
The original version of this table had Antigravity’s model at ~72% on SWE-bench Verified, trailing the other two. That was wrong. Google’s Gemini 3.1 Pro launch (February 19, 2026) reported 80.6%, a statistical tie with Opus 4.6’s 80.8%, and 68.5% on Terminal-Bench 2.0, ahead of Opus 4.6 but well behind GPT-5.3-Codex (The Decoder’s write-up has the full table). On raw model scores, the three tools were closer than their reputations. The spring-generation breakdown lives in GPT-5.4 vs Claude Opus 4.7 vs Gemini 3.1 Pro for coding, and the current-generation tier list covers Opus 5.5, GPT-6 and the rest of the November lineup.
Token efficiency is impossible to compare directly because Antigravity CLI doesn’t expose token counts. Between the other two, the famous “Codex uses 4x fewer tokens” figure came from a September 2025 Composio test on Sonnet 4 vs GPT-5. Composio’s August 2026 rerun across 47 workflows put the gap at roughly 1.2x (the Claude Code vs Codex CLI comparison walks through both tests). A lot of the remaining gap traces back to how an agent searches the repo before it acts, which is where coding agents can waste half their tokens. Antigravity CLI’s “compute effort” metric makes the same comparison impossible.
Pricing
| Plan | Antigravity CLI | Claude Code | Codex CLI |
|---|---|---|---|
| Free | $0 (Gemini 3.8 Flash, weekly starter quota) | — | Free tier (limited Codex access) |
| Budget | — | — | $8/mo (Go) |
| Entry paid | $19.99/mo (AI Pro) | $20/mo (Pro; $17 billed annually) | $20/mo (Plus) |
| Mid-tier | $100/mo (Ultra entry, 5x Pro) | $100/mo (Max, 5x Pro + $100 API credits) | $100/mo (Pro entry) |
| Top tier | $200/mo (Ultra top, 20x Pro) | $200/mo (Ultra, 10x Pro weekly + $200 API credits) | $200 and $500/mo (higher Pro tiers) |
| API fallback | Gemini API / GCP billing | $2/$10 per 1M tokens (Sonnet 5.5), $4/$20 (Opus 5.5) | Per-token OpenAI API pricing for the GPT-6 family |
At the $20/mo entry point, all three tools are nominally equivalent, and all three meter you in 5-hour windows. Claude Code’s Pro plan pairs that window with a weekly limit. Codex CLI’s Plus plan publishes estimated messages per five-hour period and notes that weekly limits “may also apply”. Antigravity CLI’s Pro tier refreshes every 5 hours under its weekly ceiling. What separates them is how well you can see the cap coming. Claude and Codex users get a usage readout. Antigravity users got lockouts the meter didn’t predict, at least in the launch-era reports.
At the $100/mo tier, the plans diverge. Claude’s Max keeps the 5-hour window but gives 5x Pro usage and adds $100 a month in API credits. Codex’s entry Pro plan has no five-hour limit at all, per OpenAI’s pricing page, which makes it the easiest of the three to run in long unattended bursts. Antigravity’s Ultra entry offers 5x Pro limits under the same weekly ceiling.
If you’re paying $200+, all three tools give you generous quotas, and Anthropic’s Ultra plan also drops the 5-hour limit for verified accounts. At that price point, predictability and transparency become the differentiators. Claude Code and Codex CLI both show you what you’re consuming. Antigravity CLI does not.
Multi-Agent Capabilities
Google’s multi-agent pitch lives mostly in the Antigravity 2.0 desktop app, where a dashboard view lets you run a refactoring agent, a test-writing agent, and a documentation agent side by side. The CLI’s share of that is handoff: a conversation started in the terminal can be exported to the desktop app and watched there. Inside the CLI itself, multi-agent work runs through built-in subagents. The changelog names pro and flash subagents, which 1.5.0 pinned to Gemini 4 Argon and Gemini 3.8 Flash after they had been silently running older models on API-key connections, plus an image-generator.
Claude Code has subagents too. The subagents guide covers the practical setup, and if you want the orchestration codified rather than driven by hand, the Claude Code workflow scripts guide builds four scripts that fan out agents deterministically. They run in parallel via the Agent tool. Claude Code’s subagent model is programmatic: you spawn agents with prompts, they return results, and you manage the orchestration in your conversation. There’s no dashboard.
Codex CLI runs parallel tasks in sandboxed environments. Each task gets its own filesystem context. It’s effective for batch operations but lacks a visual orchestration layer.
If your workflow involves multiple concurrent agents, say running a migration on three services simultaneously, Google’s desktop dashboard is the most visual way to watch them, and the CLI-to-desktop handoff is a real convenience. If you’d rather script that orchestration, a dedicated framework like the ones compared in LangGraph vs CrewAI vs OpenAI Agents SDK gives you more control. Most daily coding work is sequential anyway: read, think, write, test. The multi-agent advantage is real but narrow.
What I’d Actually Pick
The recommendations split along three axes:
Pick Claude Code if you care about code quality above everything else. Opus 5.5, now the default on every paid plan and the API, leads the current Terminal-Bench 4.0 head-to-head in Anthropic’s reporting (66.4% vs 57.9% for GPT-6 Astra), and Claude Code has won the blind comparisons published so far, though the best-known one is a dozen plan duels rather than a large study. Pricing is transparent, and Max and Ultra add monthly API credits on top of the subscription. The trade-off is higher token consumption. Claude Code thinks out loud and burns tokens doing it. If you’re on the API, Sonnet 5.5 at $2/$10 per million tokens handles most daily work at half Opus 5.5’s price.
Pick Codex CLI if you prioritize token efficiency, open-source transparency, or autonomous batch work. (Google’s own async agent, Jules, went read-only on November 11, 2026 and shuts down on December 31, so for queue-and-PR batch work the options are now Codex cloud tasks, Claude Code cloud sessions and the Copilot cloud agent.) Codex still spends fewer tokens than Claude Code, about 1.2x fewer in Composio’s latest eval. Its terminal lead from the spring (77.3% on Terminal-Bench 2.0) hasn’t carried over to the current suite, so pick it for cost and openness rather than shell accuracy. The Apache 2.0 license means you can inspect every line of code your agent runs. At $20/mo, ChatGPT Plus is the best value in this comparison. Pro starts at $100/mo with no five-hour limit, and there’s an $8/mo Go tier for lighter use.
Pick Antigravity CLI if you’re already deep in Google’s stack (GCP, Gemini API, Antigravity 2.0 desktop) and want Gemini 4 Argon in a terminal agent. The Go binary is quick, headless agy -p now covers scripted runs, and the handoff to the desktop app’s agent dashboard is a real convenience. But go in with your eyes open about the rate limits. Below the Ultra tier, the lockouts are unpredictable enough that I wouldn’t rely on it as my primary tool for deadline-sensitive work.
Skip Antigravity CLI if you’re a current Gemini CLI user who valued the open-source codebase, transparent usage metrics, and generous free tier. The replacement is a downgrade on all three of those dimensions. Google knows this. Enterprise customers get to keep Gemini CLI indefinitely. Individual developers don’t get that option.
Migration Guide for Gemini CLI Users
Gemini CLI stopped serving individual accounts on June 18, 2026. If you’re only migrating off it now, or setting up a fresh machine, here’s the practical checklist:
Staying on Google’s stack (Antigravity CLI):
# Install
curl -fsSL https://antigravity.google/cli/install.sh | bash
# Import your Gemini CLI extensions
agy plugin import gemini
# Move MCP configs (copy only the mcpServers block, then hand-edit)
cp ~/.gemini/settings.json ~/.gemini/config/mcp_config.json
# (manual edit needed — serverUrl replaces url for remote servers)
# Move skills
cp -r ~/.gemini/skills/ ~/.gemini/antigravity-cli/skills/
# Context files (GEMINI.md) work unchanged
Switching to Claude Code:
# Install (native installer; Anthropic recommends it over npm)
curl -fsSL https://claude.ai/install.sh | bash
# Create context file
# Rename your GEMINI.md to CLAUDE.md (same concept, different name)
mv GEMINI.md CLAUDE.md
# MCP config goes in .claude/settings.json
Switching to Codex CLI:
# Install
npm install -g @openai/codex
# Context file is AGENTS.md (same name Antigravity uses)
# MCP config via codex config
All three tools support MCP servers, so any MCP integrations you’ve built will carry over with minimal config changes. The main migration cost is rewriting context files and adjusting to each tool’s quirks.
FAQ
Is Antigravity CLI open source?
No. The GitHub repository contains documentation and a demo GIF, but no source code. This is a significant departure from Gemini CLI, which was fully open source under Apache 2.0. Codex CLI and OpenCode are the remaining open-source options among the major terminal agents.
Does Gemini CLI still work after June 18, 2026?
Only for enterprise users with Gemini Code Assist Standard or Enterprise licenses, or those using paid Gemini Enterprise Agent Platform API keys. Individual Pro, Ultra, and free-tier users lost access on June 18, 2026, as scheduled.
Can I use Antigravity CLI for free?
Technically yes, but only with Gemini 3.8 Flash and a small weekly starter quota; Gemini 4 Argon needs a paid plan. Launch-era reports had free users locked out for up to 7 days after 20-30 minutes of use. For anything beyond occasional one-off questions, you’ll need a paid plan.
How do Antigravity CLI rate limits compare to Claude Code and Codex CLI?
All three use 5-hour windows, and Claude and Antigravity add a weekly limit on top. Claude Pro ($20/mo) and Max ($100/mo) have both a 5-hour and a weekly limit; Ultra ($200/mo) keeps only the weekly one for verified accounts. Codex Plus ($20/mo) has a 5-hour window with weekly limits that “may also apply”, and Codex Pro (from $100/mo) currently has no five-hour limit. Antigravity refreshes every 5 hours under a weekly ceiling, with pay-as-you-go AI credits since July 2026. The difference is transparency: Claude and Codex show usage you can track and price per token on the API, while Antigravity meters an opaque “compute effort” unit.
Which terminal agent has the best code quality?
Claude Code, now defaulting to Opus 5.5. Anthropic reports Opus 5.5 at 66.4% on Terminal-Bench 4.0, ahead of the 57.9% OpenAI reported for GPT-6 Astra, and Claude Code won 8 of the 12 decided duels in Blake Crosley’s blind comparison of the two tools’ plans. On the spring models these tools launched with, the SWE-bench Verified scores were a near three-way tie: Opus 4.6 at ~81%, Gemini 3.1 Pro at 80.6%, GPT-5.4 at ~80%. Treat all of it as vendor-reported and model-level, not a measurement of the CLI harness.
Sources
- Transitioning Gemini CLI to Antigravity CLI — Google Developers Blog — official deprecation and migration announcement
- Choosing Antigravity or Gemini CLI — Google Cloud Blog — Google’s earlier (February 2026) guidance on choosing between the two
- Antigravity CLI changelog — GitHub — headless
-pmode, subagents, Gemini 4 Argon / 3.8 Flash, AI-credit errors - Antigravity CLI Deep Dive — Agentpedia — installation, current config paths, headless flags (updated October 30, 2026)
- Google Antigravity limitation — XDA Developers — launch-era rate-limit and lockout reports
- Bye-bye, Gemini CLI — The Register — community backlash and open-source concerns
- Google AI subscriptions — The Keyword — Antigravity’s compute-based quota, weekly limit, and AI-credit top-ups
- Gemini 3.1 Pro — Google and The Decoder’s coverage — February 2026 launch and benchmark table (80.6% SWE-bench Verified)
- Claude API Pricing — Anthropic — Opus 5.5 / Sonnet 5.5 per-token prices
- Claude Code model configuration — Opus 5.5 as the default on all paid plans and the API
- Using Claude Code with Pro, Max or Ultra and What is the Ultra plan? — 5-hour and weekly limits; Max 20x renamed to Ultra
- Introducing Claude Opus 5.5 — Anthropic — Terminal-Bench 4.0 figures, including the OpenAI-reported GPT-6 Astra score
- Codex pricing and Codex models — OpenAI’s current plan tiers and GPT-6 model lineup
- Composio: Claude Code vs OpenAI Codex — August 2026 47-workflow token comparison
Bottom Line
Google had something good with Gemini CLI: an open-source terminal agent with reasonable limits and a growing community. Antigravity CLI is technically better (Go is faster than Node.js, the multi-agent orchestration is real) but worse in every dimension that built trust: it’s closed-source, the rate limits are punishing, and the migration timeline was aggressive.
If you’re a Gemini CLI holdout only now setting up a replacement, the honest recommendation is: try Antigravity CLI for a day and see if the rate limits work for your usage pattern. If they don’t, Claude Code at $100/mo (Max) gives you the best code quality with capacity you can track. Codex CLI at $20/mo gives you the leanest token use and full source-code transparency. Antigravity CLI has now been in the wild for five and a half months. It has picked up a headless mode, Gemini 4 Argon, and a weekly-cap-plus-credits quota, but its usage is still metered in a unit you can’t see coming.
Claude Code remains the daily-driver pick. The rate-limit story has to change before Antigravity CLI is safe to rely on for deadline work. Whichever agent you pick, the Agent Client Protocol lets you run it in JetBrains, Zed, or Neovim through a single config file.