Gemini CLI vs Claude Code is no longer a free tool against a paid one. Google stopped serving Gemini CLI to free accounts and to Google AI Pro and Ultra subscribers on 18 June 2026, then stopped selling new Gemini Code Assist subscriptions on 9 October, so a new user can run Gemini CLI only with a paid Gemini API key or Vertex AI, billed per token. Claude Code needs a Claude plan from $20 a month or an API key.
Google's plan-based terminal agent is now Antigravity CLI (agy), free to start with higher limits on Google AI Pro ($19.99) and Ultra, so this post compares all three. Versions checked: Gemini CLI 0.66.0 (27 October 2026), Claude Code 2.1.318 (3 November 2026) and Antigravity CLI 1.5.0. Prices in US dollars, checked on 5 November 2026.
The short answer
- Pick Claude Code for the strongest published results: 64.8% on the official Terminal-Bench 4.0 leaderboard with Opus 5.5. No Gemini CLI run appears there or on Artificial Analysis's agent index.
- Keep Gemini CLI if your company already has Code Assist Standard or Enterprise seats (1,500 or 2,000 requests per user a day), or you need an Apache-2.0 agent you can fork. Those seats can't be bought new.
- Pick Antigravity CLI for Google's models at a flat price. Google AI Pro ($19.99) includes Gemini 4 Argon and, on non-trial plans, Claude Opus 5.5.
- Per token, Gemini CLI's models cost less: a month of daily agent work is $167.04 on Gemini 3.1 Pro Preview against $282.24 on Opus 5.5. A $20 plan costs less than either.

What happened to Gemini CLI
Google announced the move to Antigravity CLI on 19 May 2026. The Code Assist release notes record what followed:
| Date | What changed |
|---|---|
| 19 May 2026 | Antigravity CLI launches; Google sets the end date for consumer Gemini CLI access |
| 18 Jun 2026 | Gemini CLI stops serving free, Google AI Pro and Google AI Ultra accounts |
| 4 Sep 2026 | Code Assist no longer sold in the Cloud console to billing accounts without a subscription |
| 30 Sep 2026 | Google announces Gemini 4 Argon |
| 9 Oct 2026 | New Code Assist Standard and Enterprise subscriptions can no longer be bought |
| 27 Oct 2026 | Gemini CLI 0.66.0 adds a banner saying it isn't optimized for Gemini 4 |
| 1 Dec 2026 | Gemini 4 Argon's API price rises from $2.00 / $10.00 to $4.00 / $20.00 |
| 2027 | Existing Code Assist subscriptions stop auto-renewing |
Gemini CLI still serves existing Code Assist seats, paid Gemini API keys and Vertex AI. Its README now says "Enterprise Code Assist only", the code stays Apache-2.0, and Google promises model updates and security fixes for enterprise customers. Antigravity CLI is a separate program written in Go; its GitHub repository holds a changelog and an issue tracker but no license, so treat it as closed source.
Free tier limits: Gemini CLI vs Claude Code plans
Neither has a free tier in November 2026. Gemini CLI's free Google sign-in tier, 1,000 requests a day, ended on 18 June 2026, and Claude's Free plan has never included Claude Code. Google's free option is now Antigravity CLI.
What Gemini CLI's quota page listed, and what applies now:
| Access route | Before 18 June 2026 | Now |
|---|---|---|
| Gemini CLI, free Google sign-in | 1,000 requests a day | Not served |
| Gemini CLI, Google AI Pro / Ultra | 1,500 / 2,000 requests a day | Not served |
| Gemini CLI, unpaid Gemini API key | 250 requests a day, Flash only | Google promises access only for paid keys |
| Gemini CLI, Code Assist Standard / Enterprise | 1,500 / 2,000 requests per user a day | Same for existing seats; no new sales |
| Claude Code, Claude Free | Not included | Not included |
| Antigravity CLI, free Google account | Available from 19 May | "Basic weekly rate limits" |
Code Assist quotas count agent mode and Gemini CLI together, and one prompt can use several requests.
The paid ladders line up almost dollar for dollar (Claude pricing, Google AI plans):
| Monthly price | Claude plan (Claude Code) | Google AI plan (Antigravity CLI) |
|---|---|---|
| $0 | Free: no Claude Code | Free: weekly quota, Gemini 3.8 Flash |
| $4.99 | None | AI Plus: same Antigravity models as Free |
| $20 / $19.99 | Pro: 5-hour and weekly limits | AI Pro: 5-hour quota up to a weekly limit; Gemini 4 Argon; Claude Opus 5.5 on non-trial plans |
| $100 / $99.99 | Max: 5x Pro | Ultra (5x tier): higher Antigravity limits |
| $200 / $199.99 | Ultra: 10x Pro a week, no 5-hour limit once verified | Ultra (20x tier): highest Antigravity limits |
Google states its 5x and 20x multipliers for Gemini usage in general; for Antigravity it says only "higher" and "highest". Neither company publishes token counts, and Google's plans page says its limits depend mostly on available capacity. Claude Opus 4.6, Sonnet 4.6 and GPT-OSS 120B leave Antigravity on 18 November 2026, which leaves the free tier with Gemini 3.8 Flash. The Claude Code pricing guide covers Anthropic's limits.
Gemini CLI vs Claude Code features side by side
Antigravity CLI is included because it's where Google's development now goes.
| Gemini CLI 0.66.0 | Claude Code 2.1.318 | Antigravity CLI 1.5.0 | |
|---|---|---|---|
| Access | Existing Code Assist seats, paid API key, Vertex AI | Claude Pro and up, API key, Bedrock, Google Cloud, Foundry | Google account (Free to Ultra), Google Cloud, API key |
| License | Apache-2.0 | Proprietary, "All rights reserved" | No license published |
| Models | Gemini 3.1 Pro Preview, 3.8 Flash, 3.5 Flash-Lite, Gemma 4; Auto router | Opus 5.5 (default), Sonnet 5.5, Haiku 5.5, Fable 5.2 | Gemini 4 Argon, 3.8 Flash; Claude Opus and Sonnet 5.5 on paid plans |
| Context window | 1,048,576 tokens (Gemini 3.1 Pro, 3.8 Flash) | 1M (Opus 5.5, Sonnet 5.5) | 1,048,576 shared input and output (Argon) |
| Instructions file | GEMINI.md; context.fileName can add AGENTS.md | CLAUDE.md; reads AGENTS.md when no CLAUDE.md exists | GEMINI.md and AGENTS.md |
| MCP | gemini mcp add, settings.json | claude mcp add, .mcp.json | mcp_config.json, plugins |
| Packaging | Extensions: MCP servers, commands, hooks, subagents, skills | Plugins from marketplaces: skills, agents, hooks, MCP and LSP servers | Plugins; imports Gemini CLI extensions and Claude Code plugins |
| Sandbox | Off by default; Seatbelt, Docker, Podman, gVisor, LXC, Windows | Off by default; Seatbelt, bubblewrap; none on native Windows | Off by default, on in auto mode; Seatbelt, Linux namespaces, Windows 11 AppContainer |
| Approval modes | default, auto_edit, plan, --yolo | Manual, acceptEdits, plan, auto, dontAsk, bypassPermissions | default, accept-edits, auto (not on business accounts), plan |
| Scripting | gemini -p | claude -p | agy -p |
| GitHub | run-gemini-cli Action, @gemini-cli mentions | claude-code-action@v1, @claude mentions | No official Action found |
Context is a tie: every model here takes about a million tokens. Sandboxing is off by default in all three. Gemini CLI has the most backends; Antigravity CLI turns its sandbox on in auto mode, with no network for sandboxed commands unless a domain is allowed; Claude Code's sandbox covers shell commands only and doesn't run on native Windows.
# Gemini CLI 0.66.0: sandbox one run (or set GEMINI_SANDBOX=docker, podman, sandbox-exec, runsc or lxc)
gemini -s -p "run the test suite and fix the failures"
# Claude Code 2.1.318: inside a session
/sandbox
# Antigravity CLI 1.5.0: auto mode turns on the terminal sandbox
agy --mode=auto
The GitHub integrations differ in who pays. Claude Code's GitHub Action accepts a CLAUDE_CODE_OAUTH_TOKEN from claude setup-token, so runs draw on your Claude plan instead of API billing. The run-gemini-cli Action needs a GEMINI_API_KEY or Vertex AI; its README still promises free quotas, and its latest release, v0.1.22, is from 24 April 2026.
MCP setup is nearly identical:
gemini mcp add --transport http notion https://mcp.notion.com/mcp # writes ~/.gemini/settings.json or .gemini/settings.json
claude mcp add --transport http notion https://mcp.notion.com/mcp # local scope; --scope project writes .mcp.json
Benchmarks: what exists for each
There is no current benchmark of Gemini CLI, and Google asks for none: its README says "its harness isn't tuned for Gemini 4+ and produces unrepresentative scores." Artificial Analysis scores Antigravity CLI against Claude Code instead, with cost per task at API prices:
| Agent and model | Index | FrontierCode | Terminal-Bench 4.0 | SWE-Atlas-QnA | Cost per task | Time per task | Tokens per task |
|---|---|---|---|---|---|---|---|
| Claude Code + Opus 5.5 (max) | 62 | 56% | 63% | 66% | $11.42 | 1.0 h | 14.5M |
| Claude Code + Sonnet 5.5 (max) | 59 | 48% | 66% | 64% | $13.30 | 1.3 h | 24.1M |
| Antigravity CLI + Gemini 4 Carbon (high) | 59 | 51% | 64% | 64% | $7.74 | 29.8 min | 12.5M |
| Antigravity CLI + Gemini 4 Argon (high) | 53 | 49% | 56% | 55% | $11.73 | 34.7 min | 13.9M |

- Claude Code with Opus 5.5 leads Gemini 4 Argon by 9 index points and on every benchmark; Argon finishes in about half the time.
- Artificial Analysis prices Argon at $4.00 / $20.00, Google's rate from 1 December. At the launch price, through 30 November, a task costs about $5.87.
- Gemini 4 Carbon nearly matches Opus 5.5 at two thirds of the cost, but on 5 November it wasn't on Antigravity's model list or Google's API price list.
The official Terminal-Bench 4.0 leaderboard has Claude Code with Opus 5.5 at 64.8% and no Antigravity or Argon entry; its best Gemini result is 3.8 Flash in the neutral mini-SWE-agent harness, at 19.1%. Google's own launch figure for Argon is 77.9% on DeepSWE v1.1. For OpenAI's agent on the same index, see Codex vs Claude Code.
Monthly cost for the same usage
Price the same work per token first. Both months use the token profile from the Claude Code pricing guide: daily is 20 days of 300 model calls with a 60,000-token context and 600 output tokens; heavy is 22 days of 1,200 calls, 90,000 tokens and 700 output. In both, 92% of the context is read from cache.
| Model | Runs in | Daily month | Heavy month |
|---|---|---|---|
| Claude Opus 5.5 | Claude Code (default) | $282.24 | $1,757.18 |
| Claude Sonnet 5.5 | Claude Code | $141.12 | $878.59 |
| Gemini 3.1 Pro Preview | Gemini CLI | $167.04 | $1,039.10 |
| Gemini 3.8 Flash, to 31 Dec | Gemini CLI, Antigravity CLI | $59.94 | $375.80 |
| Gemini 4 Argon, from 1 Dec | Antigravity CLI | $253.44 | $1,567.10 |
| Gemini 4 Argon, to 30 Nov | Antigravity CLI | $126.72 | $783.55 |
Claude bills the uncached 8% as five-minute cache writes at 1.25 times input; Gemini's implicit caching has no write charge. Gemini 3.8 Flash doubles to $1.50 / $7.50 on 1 January 2027, a $119.88 daily month. Prices: Anthropic, Google AI Studio.

What the numbers mean:
- Gemini CLI's Auto mode routes between a Pro and a Flash model, so a daily month on 3.1 Pro and 3.8 Flash lands between $59.94 and $167.04: less than Opus 5.5, more than a $20 Claude Pro or Google AI Pro plan whose limits hold your work.
- A heavy month has no flat-rate Gemini CLI option left to buy. On Gemini 3.1 Pro it's $1,039.10, while Claude Ultra ($200) and Google AI Ultra ($199.99) cap the bill for their own agents.
- Measured per task, the gap closes. At Artificial Analysis's costs, 110 hard tasks cost $1,256.20 in Claude Code with Opus 5.5 and $1,290.30 in Antigravity CLI with Argon from December ($645.15 at the launch price).
Anthropic's own average is about $13 per developer per active day at API prices, close to the daily month. For 10 developers billed monthly, Claude Team standard seats cost $250; existing Code Assist Standard seats about $228, from Google's hourly license price of $0.031232877 over 730 hours; Antigravity through Google Cloud starts at $30 a seat.
To rerun the table with your own numbers:
# US$ per 1M tokens: (uncached input, cache read, output), list prices on 5 Nov 2026.
# Claude's uncached input is a 5-minute cache write (1.25x input); Gemini's implicit cache has no write charge.
PRICES = {
"Claude Opus 5.5": (5.00, 0.20, 20.00),
"Claude Sonnet 5.5": (2.50, 0.10, 10.00),
"Gemini 4 Argon (to 30 Nov)": (2.00, 0.10, 10.00),
"Gemini 4 Argon (from 1 Dec)": (4.00, 0.20, 20.00),
"Gemini 3.1 Pro Preview": (2.00, 0.20, 12.00),
"Gemini 3.8 Flash (to 31 Dec)": (0.75, 0.075, 3.75),
}
def monthly_cost(model, days, calls_per_day, context_tokens, output_tokens, cache_hit=0.92):
uncached, cached, out = PRICES[model]
calls = days * calls_per_day
per_call = (context_tokens * cache_hit * cached
+ context_tokens * (1 - cache_hit) * uncached
+ output_tokens * out) / 1e6
return round(calls * per_call, 2)
for name in PRICES:
daily = monthly_cost(name, 20, 300, 60_000, 600)
heavy = monthly_cost(name, 22, 1_200, 90_000, 700)
print(f"{name:30} daily ${daily:>9,.2f} heavy ${heavy:>9,.2f}")
# Claude Opus 5.5 daily $ 282.24 heavy $ 1,757.18
# Claude Sonnet 5.5 daily $ 141.12 heavy $ 878.59
# Gemini 4 Argon (to 30 Nov) daily $ 126.72 heavy $ 783.55
# Gemini 4 Argon (from 1 Dec) daily $ 253.44 heavy $ 1,567.10
# Gemini 3.1 Pro Preview daily $ 167.04 heavy $ 1,039.10
# Gemini 3.8 Flash (to 31 Dec) daily $ 59.94 heavy $ 375.80
The Gemini API pricing guide covers batch and long-context rates.
Claude Code vs Gemini CLI: who should pick which
- Claude Code, for the strongest published results on long tasks at a flat monthly price. Pro at $20 covers most single-developer days; Max and Ultra cover parallel agents.
- Gemini CLI, if your organization already holds Code Assist seats (until renewal ends in 2027), or you need an open-source agent you can audit, fork or run on Vertex AI.
- Gemini CLI on a paid API key, for per-token billing on Gemini 3.1 Pro or 3.8 Flash, such as CI jobs under a project spend cap. For daily interactive work, a plan costs less.
- Antigravity CLI on Google AI Pro, for Gemini 4 Argon at $19.99, or one subscription that also reaches Claude Opus 5.5 inside Google's agent.
- Either Google CLI, if you work on native Windows and want commands sandboxed (Antigravity's sandbox needs Windows 11 24H2 or later). Claude Code needs WSL2 for its sandbox.
For agents beyond these three, see Claude Code alternatives.
Moving off Gemini CLI
Start with the instructions file. Rename GEMINI.md to AGENTS.md: Claude Code reads it when a repository has no CLAUDE.md, Antigravity CLI reads it natively, and Gemini CLI keeps reading it with one setting in .gemini/settings.json:
{
"context": {
"fileName": ["AGENTS.md", "GEMINI.md"]
}
}
For Antigravity CLI, Google's migration guide imports extensions as plugins and custom commands as skills:
curl -fsSL https://antigravity.google/cli/install.sh | bash # Windows: irm https://antigravity.google/cli/install.ps1 | iex
agy plugin import gemini
MCP servers move to ~/.gemini/config/mcp_config.json, with url or httpUrl renamed serverUrl; workspace skills move from .gemini/skills/ to .agents/skills/ by hand.
For Claude Code, install with curl -fsSL https://claude.ai/install.sh | bash and re-add each server from gemini mcp list with claude mcp add. The Claude Code install guide covers Windows and sign-in.
What developers report
- Google's announcement drew 376 reactions on GitHub, 301 of them thumbs-down (GitHub discussion #27274, 19 May 2026, link).
- "Gemini has the most generous limits on the $20 plan." (a Hacker News commenter comparing Antigravity with Argon, Codex and Claude Code, 28 Oct 2026, link)
- "The quotas on Pro are about 2 hours of usage per week for Argon." (a Hacker News commenter using Antigravity CLI, 29 Oct 2026, link)
- A Google AI Plus subscriber said Antigravity with Gemini 3.8 Flash spent a week's quota in about 10 minutes; similar prompts get 2 to 3 hours of work on $20 Claude and ChatGPT plans (Hacker News, 29 Oct 2026, link).
- "Gemini CLI was disastrous. The Antigravity CLI is much better." (a Hacker News commenter, 13 Oct 2026, link)
- A Hacker News commenter called Antigravity CLI "unworkable" because auto mode still stopped for approvals (14 Oct 2026, link); the 1.5.0 changelog says auto mode now asks less often on safe commands.
The quota reports conflict partly because they compare Argon on AI Pro with 3.8 Flash on AI Plus.
Where Aight fits
Neither CLI can point at Aight. Gemini CLI speaks Google's Gemini API, Claude Code speaks Anthropic's Messages API, and Aight's API reference documents an OpenAI-compatible Chat Completions API. For the agent itself, a plan costs less: Aight charges list price for Opus 5.5 ($4.00 / $20.00), so the heavy month above would cost about as much there as at Anthropic, against $200 for Claude Ultra.
Aight fits the second-opinion step many developers already run by hand: sending Claude's diff to a Gemini model, or the reverse, from a script or CI job. One sk-aight- key reaches Gemini 3.1 Pro Preview at Google's price ($2.00 / $12.00) and Claude Opus 5.5, with a monthly cap per key and a list of allowed models. Configured as an OpenAI-compatible provider with the openai Python SDK (3.45.0):
import os
import subprocess
from openai import OpenAI
client = OpenAI(
base_url="https://api.getaight.ai/v1",
api_key=os.environ["AIGHT_API_KEY"],
)
diff = subprocess.run(
["git", "diff", "main...HEAD"], capture_output=True, text=True, check=True
).stdout
resp = client.chat.completions.create(
model="gemini-3.1-pro-preview", # or "claude-opus-5-5" to review Gemini's work
messages=[{
"role": "user",
"content": "Review this diff. List bugs, missing tests and risky changes, "
"each with the file and line.\n\n" + diff,
}],
)
print(resp.choices[0].message.content)
FAQ
Is Gemini CLI still free?
Not for new users. Since 18 June 2026 it doesn't serve free accounts or Google AI Pro and Ultra subscribers, and Google promises access only through paid Gemini API keys and Vertex AI. The quota page still lists 250 requests a day for an unpaid key, but Google hasn't said that route is supported. Google's free terminal agent is now Antigravity CLI.
Is Gemini CLI better than Claude Code?
For most new users, no. Claude Code has current benchmark results and a $20 entry plan, while Gemini CLI is limited to enterprise seats and per-token keys, and Google says it isn't tuned for Gemini 4. Gemini CLI still wins on license: it's Apache-2.0, and Claude Code is proprietary.
Can Claude Code use Gemini models?
Not with Anthropic's support. Its gateway docs say Anthropic doesn't support routing Claude Code to non-Claude models. The reverse works officially: Antigravity CLI runs Claude Opus 5.5 and Sonnet 5.5 on Google AI Pro (non-trial) and Ultra plans.
Does Gemini CLI work with Gemini 4 Argon?
Not officially. Gemini CLI 0.66.0's model list stops at Gemini 3.1 Pro Preview and 3.8 Flash, and its docs say to use Antigravity for Argon. In Antigravity CLI, Argon needs a paid Google AI Pro plan, Ultra, Google Cloud, or a paid API key.
Sources
- Google, Gemini CLI: README, releases, @google/gemini-cli on npm, source at v0.66.0 (models.ts, settingsSchema.ts, docs/tools/mcp-server.md, docs/cli/gemini-md.md), discussions #27274 and #28017
- Gemini CLI docs: Quotas and pricing, Authentication, Model selection, Sandboxing, Extensions, Changelogs
- Google for Developers Blog, Transitioning Gemini CLI to Antigravity CLI, 19 May 2026
- Google Cloud: Gemini Code Assist release notes, Quotas and limits, Gemini for Google Cloud pricing
- Google Antigravity: Antigravity CLI, Pricing, Plans, Models, Quickstart, Authenticate, Terminal sandbox, Gemini CLI migration, antigravity-cli on GitHub (README and CHANGELOG)
- Google AI plans: gemini.google/subscriptions, Google One AI plans
- Google AI Studio: Gemini API pricing, Gemini 4 Argon model page, Gemini 3.8 Flash, Gemini 3.1 Pro; Google, Gemini 4 Argon launch post
- GitHub Actions: google-github-actions/run-gemini-cli and its releases; Claude Code docs, GitHub Actions
- Anthropic: Claude pricing, API pricing, claude-code LICENSE, @anthropic-ai/claude-code on npm
- Claude Code docs: Overview, Model configuration, Sandboxing, Permission modes, MCP, Plugins, Memory and AGENTS.md, CLI reference, Costs, LLM gateways
- Benchmarks: Artificial Analysis, Antigravity CLI vs Claude Code, Artificial Analysis, Gemini 4 Argon, Terminal-Bench 4.0 leaderboard
- Hacker News comments: 50278012, 50288698, 50284169, 50079370, 50086246
- OpenAI Python SDK: openai on PyPI
- Aight: Models and pricing, API reference
*Checked on 5 November 2026. Updated 17 November 2026: Fable 5.2 is Claude Code's fable model; Claude Code 2.1.326 changes nothing else compared here.*