Gemini CLI vs Claude Code (2026): Models, Pricing and Which to Choose

Key Takeaways

• Since June 18, 2026, Gemini CLI serves Code Assist customers and paid Gemini API or Vertex AI users.
• Free and Google AI Pro or Ultra users moved to Antigravity CLI, open to all since May 19, 2026.
• Claude Code defaults to Opus 5.5, with a 1M-token context window and 128K output.
• Both tools support MCP, subagents, plan mode and a project file (GEMINI.md or CLAUDE.md).
• Claude Code starts at $20 a month, while Gemini CLI needs a Code Assist license or paid API key.
• Decide by license first, then by accepted changes, review time and cost on your own repository.

A Platform Team Choosing a Terminal Agent This Month

Picture a five-person platform team on Google Cloud with no Code Assist license. It wants an agent that edits across files and runs tests.

A year ago the answer looked easy, because Gemini CLI was free with a personal Google account. Since June 18, 2026, however, it no longer serves free or AI Pro and Ultra users.

So the team now picks between Gemini CLI on a paid API key and Claude Code seats. Every version and price in this Gemini CLI vs Claude Code comparison was checked on November 3, 2026.

What Gemini CLI and Claude Code Are in November 2026

Both products are terminal coding agents that read a repository, edit files and run commands. They belong to the wider family of AI coding agents, yet they now serve very different audiences.

Gemini CLI: Google’s Open-Source Agent, Now for Code Assist Teams

Gemini CLI is open source under the Apache 2.0 license, hosted in the google-gemini/gemini-cli repository with over 100,000 GitHub stars. Since October 27, 2026, its latest stable release has been v0.66.0.

Its audience changed in 2026, though. The README now targets Code Assist Standard and Enterprise customers, plus users with paid Gemini API or Vertex AI keys.

Claude Code: Anthropic’s Agent Across Terminal, IDE, Desktop and Web

Claude Code reads your codebase, edits files and runs commands, and it asks for approval before risky steps. It runs in the terminal, in VS Code and JetBrains, in a desktop app and on the web.

The latest entry in the Claude Code releases on GitHub is v2.1.317, dated November 3, 2026. Claude Code is proprietary, so Anthropic’s Commercial Terms of Service govern its use.

What Happened to Gemini CLI on June 18, 2026?

Google announced the change on May 19, 2026, in a Google Developers Blog post. Then, on June 18, Gemini CLI and the Code Assist IDE extensions stopped serving Google AI Pro, Ultra and free individual accounts.

Those users moved to Antigravity CLI, which is built in Go and shares its agent core with Antigravity 2.0. It keeps Agent Skills, Hooks, Subagents and Extensions, now shipped as plugins. Google notes that feature parity was not complete at launch.

Licensed teams, however, kept Gemini CLI. Google also stopped selling new Code Assist subscriptions, although its Code Assist overview gives no date. Our Google Antigravity vs Claude Code comparison covers the successor tool in depth.

Gemini CLI vs Claude Code at a Glance

The two tools now differ more in access than in features. So read the license and payment rows first, since they decide whether a tool is open to you.

Table 1: Both tools side by side, checked November 3, 2026 against vendor docs and GitHub releases.

Supported Models and Context Windows

The model behind an agent shapes its output more than the shell does. For model families, see ChatGPT vs Gemini vs Claude and our list of top LLMs.

Which Gemini Models Does Gemini CLI Use?

Gemini CLI v0.66.0 works with the Gemini 3.x family. Its list includes gemini-3-pro-preview, gemini-3.1-pro-preview, gemini-3.5-flash and gemini-3.1-flash-lite. It also adds gemini-3.8-flash when the account has GA access.

Most users leave it on Auto (Gemini 3) or Auto (Gemini 2.5), which picks Pro or Flash by task. Our Google Gemini AI guide explains the family.

Google states that Gemini CLI is not optimized for Gemini 4 and later. Instead, Gemini 4 Argon, with a 1M-token context per the Gemini API model docs, runs through Antigravity CLI, Gemini CLI’s successor.

Which Claude Models Does Claude Code Use?

Claude Code defaults to Opus 5.5 on Pro, Max, Ultra, Team, Enterprise and the Anthropic API, per its model configuration docs. Amazon Bedrock and Google Cloud Agent Platform also default to Opus 5.5.

On the Anthropic API, the aliases opus, sonnet and haiku map to Opus 5.5, Sonnet 5.5 and Haiku 5.5, while fable maps to Fable 5.1. The opusplan alias uses Opus in plan mode and Sonnet for execution.

Cloud accounts need one extra check. Aliases can map to earlier versions on some providers, and Microsoft Foundry defaults to Sonnet 4.5, so confirm the model id in each session.

Does a Bigger Context Window Mean Better Code?

Opus 5.5 and Sonnet 5.5 both accept 1M tokens of context and return up to 128K tokens, so one session can hold much of a repository.

Even so, a large window does not guarantee correct edits, since the model still has to pick the right files. Claude Code auto-compacts long sessions, which is why agentic context engineering matters more than raw window size.

For Gemini CLI, the window depends on the model Auto picks, so measure how often the agent edits the right place.

Gemini CLI vs Claude Code for Real Coding Work

Both tools handle common coding jobs, so this section compares setup and review effort. For wider options, see our guide to AI code assistants.

Writing New Code

New code is the easiest job to hand an agent, and also the easiest to merge without reading. Say you ask each tool for a new REST endpoint with tests.

Gemini CLI on Auto may route a short prompt like that to a Flash model, which suits quick prototypes on Google Cloud. Claude Code answers from Opus 5.5 by default, in the terminal, an IDE, the desktop app or the web.

Either way, read the endpoint and its tests before you merge. For throwaway experiments, vibe coding habits apply to both, so keep that review step even when the code looks right.

Refactoring Across Many Files

Multi-file refactors are where planning matters most, because one wrong file can break the build. Both tools have a plan mode, so the agent lists the files it will touch before it edits.

Claude Code subagents can split a refactor into jobs such as fixing call sites and tests. Both tools can run parallel attempts in git worktrees, so each stays on its own branch.

For example, a three-file rename with one failing test follows the same steps in either tool.

1. Ask for a plan first. The agent lists the three files, the renamed symbol and the test it expects to break.
2. Approve the plan, and then let the agent edit. Each change appears as a diff.
3. Run the test suite, then let the agent read the failure and propose a fix.
4. Review the final diff before you commit, and ignore the chat summary.

Debugging and Failing Tests

Both agents run a test command, read the stack trace and propose a patch. The safer pattern makes the test runner the judge, so a change counts only when the suite passes.

Both can enforce that with hooks, and both keep checkpoints you can rewind if a fix backfires. Our guide to generative AI in software testing covers that test layer.

Working in Large Repositories

Large repositories need persistent instructions, because an agent cannot guess your conventions. So Gemini CLI reads GEMINI.md and Claude Code reads CLAUDE.md, each holding build commands and style rules.

Take a monorepo where a payments service depends on a shared library two folders away. Put the library’s build command and test rules in the instruction file, so the agent edits the service without breaking the library.

Gemini CLI also accepts –include-directories, which adds folders outside the working directory. Claude Code, meanwhile, leans on the 1M-token context of its default model.

Agents, Subagents, Plan Mode and Recovery

Agent features decide how much work you can hand off without watching every step. Both tools ship the same core ideas, though they package them differently.

These patterns come from AI agent orchestration and the engineering layer around an agent loop. The same ideas also drive LLM-powered autonomous agents outside coding. The episode below explains why the Hermes agent drew so much attention.

In Claude Code, subagents are Markdown files with YAML frontmatter that define specialized assistants. Hooks run your scripts at set points, while agent teams remain experimental and disabled by default as of November 2026.

Gemini CLI also has plan mode, hooks, checkpointing, rewind, git worktrees and subagents. One detail trips people up, because the /model command does not override the model a sub-agent uses.

What Happens When a Long Task Goes Wrong?

Long tasks fail when an early assumption is wrong or a command changes too much. Gemini CLI saves checkpoints, so rewind restores files to an earlier point.

Claude Code also checkpoints each prompt, so /rewind restores earlier files, and its hooks can block an action that breaks policy. In both tools, commit before each run so git gives you a way back.

Task size matters too. A short task with a clear test is far easier to undo than one long session that touches every folder.

MCP, Extensions, Skills and IDE Integration

MCP, the Model Context Protocol, lets an agent call outside tools such as issue trackers and databases. Both tools support it, so packaging is what differs. Our guide to MCP for context-aware AI agents explains the protocol.

Claude Code connects to hundreds of tools and data sources through MCP, per its MCP documentation. Gemini CLI configures MCP servers in its settings, and its extensions bundle commands with MCP servers into one install.

Both tools also support skills for repeatable tasks. If your agents will talk to other agents, the MCP vs A2A comparison shows where each protocol fits.

IDE support, however, differs more. Claude Code has VS Code and JetBrains extensions plus desktop and web apps, while Gemini CLI pairs with a VS Code companion.

When an MCP Config Fails, and the Fix

Most MCP failures come from small config mistakes, not from the protocol itself. Claude Code’s MCP docs describe one that catches people who paste server entries by hand.

An entry with a url but no type looks harmless. Claude Code, however, reads an entry with no type as a local stdio server, so it skips this one.

// .mcp.json that fails: url but no type
{
  "mcpServers": {
    "notion": {
      "url": "https://mcp.notion.com/mcp"
    }
  }
}

// Claude Code skips the server and reports:
// MCP server "notion" has a "url" but no "type"; add "type": "http" (or "sse" / "ws") to this entry

// The fix: declare the transport
{
  "mcpServers": {
    "notion": {
      "type": "http",
      "url": "https://mcp.notion.com/mcp"
    }
  }
}

A second failure is a safety feature. In a remote server’s url and headers, Claude Code reads credential variables such as ANTHROPIC_API_KEY as empty. The server then usually rejects the call with a 401.

That block stops a shared .mcp.json from sending your own credentials to a server it names. So give each server its own token, and never reuse your agent’s key.

Every MCP server also widens what an agent can reach, so grant access per server and keep write tools behind approval. That is the same least-privilege rule security teams apply to people.

How to Install and Set Up Each Tool

Setup takes a few minutes for either tool, so the main difference is sign-in. Gemini CLI needs a license or a paid key, while Claude Code needs a paid Claude plan or an API key.

Installing Gemini CLI

# run once without installing
npx @google/gemini-cli

# or install it globally
npm install -g @google/gemini-cli
brew install gemini-cli

Then pick one sign-in path. Set GOOGLE_CLOUD_PROJECT for a Code Assist license, GEMINI_API_KEY for a paid API key, or the Vertex AI variables for a Google Cloud project.

export GOOGLE_CLOUD_PROJECT="your-project-id"
# or
export GEMINI_API_KEY="your-paid-api-key"

# headless run with JSON output for CI
gemini -p "Summarize the failing tests" --output-format json

Installing Claude Code

curl -fsSL https://claude.ai/install.sh | bash

Windows uses the install.cmd script, and Homebrew also works. After the install, sign in with a paid plan or an API key.

Claude Code also runs headless in CI, so a review task can run on every pull request. Check each CLI reference before scripting, since flags change.

Gemini CLI vs Claude Code Pricing and Access in November 2026

Pricing works differently for each tool. Claude Code sells plans directly, while Gemini CLI rides on a Google license or on metered API usage.

Is Gemini CLI Still Free?

Gemini CLI is no longer free for new individuals, since the free and AI Pro or Ultra paths ended on June 18, 2026.

Licensed teams get a daily quota instead. Code Assist Standard allows 1,500 model requests per user per day, and Enterprise allows 2,000, per the Gemini CLI quota page. Usage on a paid API key varies by tier.

New Code Assist subscriptions are no longer for sale, so a team without a license needs a paid Gemini API or Vertex AI key.

What Claude Code Costs

Claude Code comes with paid Claude plans, but the free plan does not include it. The table lists Claude’s plan prices as of November 3, 2026.

Table 2: Claude plans and Claude Code access, November 2026.

API Rates Behind Both Tools

API rates matter when you run either tool on a key, so the table lists prices per 1M tokens as of November 3, 2026.

Table 3: API rates for the models behind each tool.

Once several teams share keys, an LLM gateway logs spend by team and model in one place.

How Should You Compare Total Cost?

Seat price and token price both miss the real number, which is cost per accepted change. That figure adds tokens, retries, CI minutes and review time, then divides by the changes that merge.

So log the model id, usage, retries, CI minutes and reviewer minutes per change during a pilot. After all, a cheap run that needs heavy review can cost more than a pricier run that merges cleanly.

Security, Licensing and Governance

A coding agent runs commands on real machines, so its permission model matters as much as its output. Both tools gate risky actions, but they give security teams different controls.

Gemini CLI has a sandbox, trusted folders, hooks, telemetry settings and an enterprise guide for admins. Claude Code uses permission prompts, an optional OS-level sandbox for shell commands, and hooks that act as policy checks. Admins can also restrict models per organization.

Both tools share one major risk, prompt injection. A hostile instruction can hide in a repository file, an issue or an MCP response, so limit which servers an agent can reach. Our guides to agentic AI security and agentic AI governance cover these controls.

Open Source vs Proprietary: Does the License Matter?

The license matters most when auditors get involved. Because Gemini CLI’s code is open under Apache 2.0, a security team can read and fork the agent itself.

Claude Code is proprietary, so teams rely on its permission settings and logs instead of source review. Teams with strict source-audit rules may prefer Gemini CLI for that reason.

How to Benchmark Gemini CLI vs Claude Code Fairly

Public benchmarks measure models, while your team needs to know how a tool performs on its own code.

Google itself warns against using Gemini CLI to benchmark Gemini models. Meanwhile, Anthropic reports Opus 5.5 at 66.4% on Terminal-Bench 4.0 at xhigh effort, against 55.8% for Fable 5.1. Opus 5.5 matches or beats Fable 5.1 on most listed benchmarks, yet that is a model score with no Gemini CLI comparison.

A fair pilot tests both tools under the same conditions, in six steps.

1. Pin one commit of one repository with good test coverage.
2. Write a fixed task list of bug fixes and refactors.
3. Give both tools the same permissions and instructions.
4. Record the CLI version and model id for every run.
5. Measure accepted changes, reviewer minutes and spend per task.
6. Repeat each task at least twice, because agent runs vary.

Then score the results with one rubric for both tools. Our guides to AI agent evaluation and an LLM evaluation framework show how to build that rubric.

Limitations of Each Tool

Each tool’s gaps hurt a different kind of team, so match these lists against your own situation before you commit.

Gemini CLI Limitations

Most of Gemini CLI’s gaps are about access, not capability. They hurt individuals and new teams most, because a license or a paid key now comes first.

1. It has been closed to new individual users since June 18, 2026.
2. It is not tuned for Gemini 4.
3. New Code Assist subscriptions are no longer sold.
4. Auto routing can switch models between runs.

Claude Code Limitations

Claude Code’s gaps are about cost control and source visibility. They hurt teams on tight budgets and teams whose auditors want to read the agent’s code.

1. No free plan includes Claude Code.
2. The code is proprietary, so you cannot audit its source.
3. Subscriptions carry usage limits per 5-hour and weekly window.
4. Alias versions differ by cloud provider.
5. Agent teams are still experimental.

Gemini CLI vs Claude Code: Which Should You Choose?

The right choice depends on what your team already owns, so start with license, then workload, then measured results.

Choose Gemini CLI If…

• You hold an active Code Assist Standard or Enterprise license.
• Your stack runs on Google Cloud and Vertex AI.
• Your security team needs open-source code it can audit.

Choose Claude Code If…

• You are starting fresh without a Code Assist license.
• Multi-file refactors and code review fill most of your week.
• You want one agent across terminal, IDE, desktop and web sessions.

If You Need Gemini 4: Gemini CLI’s Successor

If Gemini 4 Argon is a must, the Gemini side of this choice now continues in Antigravity CLI, Gemini CLI’s successor. Gemini CLI does not target that model, so read our head-to-head of Antigravity and Claude Code before you switch.

Can You Use Both?

Yes. Route cloud infrastructure tasks to Gemini CLI under the license, and send refactors and reviews to Claude Code. Each tool then works where its license or strengths fit best.

Also keep one conventions source and generate GEMINI.md and CLAUDE.md from it, so both agents follow the same rules.

How Kanerika Rolls Out AI Coding Agents for Engineering Teams

Kanerika is certified in the Claude Partner Network (Services Track), announced in October 2026, so weigh this section with that in mind. The rollout method below applies to Gemini CLI and Claude Code alike.

Kanerika’s agentic AI services team follows five stages when a client adopts a coding agent.

1. Pick one pilot repository with good test coverage and a pinned commit.
2. Write the permission policy first, covering allowed commands, an MCP allowlist, secrets kept out of reach, and sandbox settings.
3. Put a test gate in CI, so no agent-written change merges without passing the tests a human change must pass.
4. Run the same task list through each tool, and log CLI version, model id, accepted changes, review minutes and spend.
5. Decide on cost per accepted change, then roll out with shared CLAUDE.md or GEMINI.md conventions and a review owner per repository.

Stages 1 and 2 keep the pilot small and safe, because scope limits damage. A well-tested repository gives clear signals, while a written policy stops the agent from reaching systems nobody approved.

Why the Test Gate Matters Most

Stage 3 carries the most weight, because it decides what merges. Kanerika built that kind of gate for a digital construction platform, using a unified BDD test framework with Serenity BDD and Cucumber.

The team also wrote reusable test code and ran automated tests with real-time reporting in the CI pipeline. As a result, pipeline execution time fell by 60%, and product releases became 90% faster.

FLIP, Kanerika’s automated conversion platform, follows the same generate-then-validate idea at migration scale. It converts legacy pipeline logic and then validates the output automatically, so engineers review exceptions instead of every line. A CI test gate does the same job for agent-written code, since nothing merges until the checks pass.

Habits That Keep an Agent Pilot Safe

A few habits make any pilot safer. Start agents in plan mode on unfamiliar repositories, keep one instruction file per repository, and review diffs rather than chat summaries. Each habit keeps a person in charge of what merges.

Also cap spend per developer during the pilot. For custom agents beyond the CLI tools, Kanerika’s AI application development team designs and tests the full application. Each repository then keeps its own acceptance and spend numbers, so leaders can see whether Gemini CLI or Claude Code pays off.

More AI Coding Agent Comparisons

Gemini CLI and Claude Code are two of several terminal agents worth a pilot. See GPT Codex vs Claude Code, GitHub Copilot vs Claude Code vs Cursor vs Windsurf, OpenClaw vs Claude Code vs Claude Cowork and Grok Build vs Hermes vs Claude Code.

Bottom Line: Gemini CLI or Claude Code?

Keep Gemini CLI if you already hold a Code Assist license or need source you can audit. Pick Claude Code for a new team, since it needs no Google license and starts at $20 a month.

For the wider field, our guide to AI coding agents maps the category. Our Codex comparison and Copilot, Cursor and Windsurf comparison test Claude Code against other rivals.

Then pilot both on one repository, and let cost per accepted change settle Gemini CLI vs Claude Code for your team.

Frequently Asked Questions