IT Pro Tutorials Claude Code vs Gemini CLI vs Codex CLI: Which AI Coding CLI Wins in 2026?
Three AI labs. Three terminal agents. Three different philosophies. Claude Code bets on autonomous correctness, Gemini CLI leads with a free tier and a million-token context window, and Codex CLI prioritizes sandbox safety. But which one actually delivers when you need to write infrastructure code, debug a broken Dockerfile, or generate an Ansible playbook?
We put all three through the same five tasks to find out. No cherry-picked demos. No marketing benchmarks. Just real infrastructure work, timed and measured.
The Contenders at a Glance
| Claude Code | Gemini CLI | Codex CLI | |
|---|---|---|---|
| Provider | Anthropic | OpenAI | |
| Model | Opus 4.6 / Sonnet 4.6 | Gemini 2.5 Flash / Pro | GPT-5.3 Codex / codex-mini |
| Context Window | 200K tokens | 1M tokens | 192K tokens |
| Free Tier | No ($20/mo Pro) | Yes (1,000 req/day) | No ($20/mo Plus) |
| Install | npm i -g @anthropic-ai/claude-code |
npx https://github.com/google-gemini/gemini-cli |
npm i -g @openai/codex |
| OS | macOS, Linux, Windows (WSL) | macOS, Linux, Windows | macOS, Linux |
| Sandbox | No (full system access) | No (full system access) | Yes (sandboxed execution) |
| Open Source | No | Yes | Yes |
Test Setup
We ran each tool on five infrastructure tasks that reflect real sysadmin and DevOps work. Each task started from the same clean Git repo with identical base files.
Test environment: Ubuntu 22.04, 4 vCPUs, 8 GB RAM, Node.js 20, Python 3.11
Tasks:
1. Generate a Docker Compose stack — Nginx reverse proxy with SSL, two backend services, health checks, and a shared network
2. Write an Ansible playbook — Install and configure Docker on Ubuntu 22.04 with security hardening
3. Debug a broken Dockerfile — Find and fix three intentional errors in a multi-stage Node.js build
4. Create a Bash monitoring script — CPU, RAM, disk usage with threshold alerts and logging
5. Refactor a Python config parser — Break a 400-line single-file script into modules with proper error handling
Each tool got the same prompt. We measured time to completion, token usage, and whether the output worked on first run.
Results: Task by Task
Task 1: Docker Compose Generation
| Metric | Claude Code | Gemini CLI | Codex CLI |
|---|---|---|---|
| Time | 45 seconds | 38 seconds | 1 min 52 sec |
| Worked first run? | Yes | Yes (missing health check) | Yes |
| Quality | Complete with comments | Functional, minimal comments | Complete, over-commented |
All three produced working Docker Compose files. Claude Code included health checks, restart policies, and inline comments explaining each choice. Gemini CLI was fastest but missed the health check on one service. Codex CLI was slowest but thorough, adding a .env file and a README unprompted.
Winner: Claude Code (completeness) / Gemini CLI (speed)
Task 2: Ansible Playbook
| Metric | Claude Code | Gemini CLI | Codex CLI |
|---|---|---|---|
| Time | 1 min 12 sec | 58 seconds | 2 min 45 sec |
| Worked first run? | Yes | Partial (missing handler) | Yes |
| Quality | Production-ready | Needs minor fixes | Production-ready |
Claude Code generated a complete playbook with roles, handlers, and idempotent tasks. It even added a molecule test scaffolding without being asked. Gemini CLI produced a working playbook but forgot a handler for restarting Docker after config changes. Codex CLI was thorough but slow, producing a well-structured playbook with tags and variables.
Gemini CLI’s Google Search grounding pulled in the latest Docker APT repository URL, which was a nice touch. Claude Code used a slightly older URL that still works. Small detail, but it shows Search grounding’s practical value.
Winner: Claude Code (quality) / Gemini CLI (current docs)
Task 3: Dockerfile Debugging
| Metric | Claude Code | Gemini CLI | Codex CLI |
|---|---|---|---|
| Time | 32 seconds | 1 min 5 sec | 1 min 38 sec |
| Found all 3 bugs? | Yes | 2 of 3 | Yes |
| Explanation quality | Excellent | Good | Excellent |
This is where Claude Code’s reasoning shines. It identified all three bugs immediately, explained the root cause of each, and applied fixes in a single pass. Codex CLI found all three but took longer, methodically testing each fix in its sandbox. Gemini CLI missed a subtle COPY --from stage reference error, finding it only after a second prompt.
Winner: Claude Code (clear leader)
Task 4: Bash Monitoring Script
| Metric | Claude Code | Gemini CLI | Codex CLI |
|---|---|---|---|
| Time | 55 seconds | 42 seconds | 2 min 10 sec |
| Worked first run? | Yes | Yes | Yes (in sandbox) |
| Quality | Full-featured | Clean and functional | Over-engineered |
Gemini CLI produced the cleanest script: concise, well-structured, and exactly what was asked for. Claude Code added features beyond the spec (log rotation, email alerting, systemd timer integration) which could be helpful or noisy depending on your perspective. Codex CLI built an entire monitoring framework with config files, which was impressive but far more than needed.
Winner: Gemini CLI (pragmatism) / Claude Code (features)
Task 5: Python Refactoring
| Metric | Claude Code | Gemini CLI | Codex CLI |
|---|---|---|---|
| Time | 2 min 8 sec | 3 min 42 sec | 4 min 15 sec |
| Worked first run? | Yes | No (import error) | Yes |
| Files created | 5 | 4 | 6 |
Multi-file refactoring is where the gap between tools becomes clear. Claude Code restructured the module cleanly, maintained backward compatibility, and updated all imports across files. Gemini CLI struggled with circular imports on the first attempt and needed a follow-up prompt. Codex CLI produced a solid result but took the longest, carefully testing each module in its sandbox.
Winner: Claude Code (decisive lead)
Overall Scorecard
| Category | Claude Code | Gemini CLI | Codex CLI |
|---|---|---|---|
| Speed | Fast | Fastest (simple tasks) | Slowest |
| Accuracy | Highest | Good (occasional misses) | High |
| Multi-file tasks | Excellent | Weak | Good |
| Simple scripts | Excellent | Excellent | Over-engineers |
| Cost per test run | ~$4.80 | $0 (free tier) | ~$3.50 |
| Safety | Manual review | Manual review | Sandboxed |
| Overall Score | 9/10 | 7/10 | 7.5/10 |
Cost Comparison
This is where Gemini CLI’s free tier changes the equation entirely.
| Tool | 5-Task Benchmark Cost | Estimated Monthly Cost (daily use) |
|---|---|---|
| Claude Code | $4.80 | $50-100 |
| Gemini CLI | $0 (free tier) | $0 (if under 1,000 req/day) |
| Codex CLI | $3.50 | $40-80 (+ $20 ChatGPT Plus) |
Claude Code costs more but saves time. If your time is worth $50/hour and Claude Code saves 30 minutes per day over Gemini CLI, the math works out. But if you are a solo sysadmin managing a small fleet and need quick scripts, Gemini CLI’s free tier is hard to argue against.
When to Use Each
Choose Claude Code When:
- You need multi-file refactoring or complex autonomous tasks
- Accuracy matters more than cost (production infrastructure)
- You work with large codebases that need deep understanding
- You want the least amount of hand-holding and re-prompting
- You manage Ansible playbooks and need production-ready output
Choose Gemini CLI When:
- Budget is the primary constraint
- You need quick, one-off scripts and configs
- You want the largest context window for big projects
- You value Google Search grounding for current documentation
- You are evaluating AI CLI tools for the first time
- You work with cloud services and need free tooling
Choose Codex CLI When:
- Safety is your top priority (the sandbox prevents destructive actions)
- You already pay for ChatGPT Plus and want terminal access
- You need careful, methodical code generation over speed
- Compliance requires sandboxed execution
- You prefer reviewing AI work before it touches your filesystem
What About Aider and Cline CLI?
This comparison focused on the “big three” from major AI labs, but open-source alternatives deserve mention.
Aider brings model flexibility: use Claude, GPT, Gemini, or local models through a single tool. Its Git integration is the best in the category. If you want Claude Code’s quality with Gemini CLI’s cost, Aider with a Claude API key is a compelling middle ground.
Cline CLI adds approve-everything safety (like Codex’s sandbox but without the restrictions), parallel agents, and headless CI/CD mode. It works with any model provider. If you want maximum control over what the AI does, Cline is worth evaluating.
For a full overview of all options, see our complete guide to AI coding CLI tools.
Verdict
Claude Code wins on capability. It is the most reliable, most accurate, and fastest tool for complex infrastructure tasks. If you write Ansible, Terraform, or Docker configs daily, it pays for itself in time saved.
Gemini CLI wins on value. The free tier is real and useful, not a demo. For quick scripts, one-off configs, and learning, you cannot beat free with a 1M token context window.
Codex CLI wins on safety. The sandbox is not a gimmick. If you need guardrails, Codex delivers them without sacrificing too much capability.
Our recommendation for most IT pros: start with Gemini CLI (free) to build the habit, then add Claude Code when you hit tasks that need more horsepower. Keep both installed. Use the right tool for the job.
If you are already using n8n for automation workflows, pairing it with an AI coding CLI for script generation is a natural next step in your automation toolkit.
