Gemini CLI wins on price (free tier with 1,000 requests/day) and context window (1M tokens standard). Claude Code wins on code accuracy (80.9% SWE-bench, 95% first-pass accuracy) and multi-agent orchestration via Agent Teams. Codex CLI wins on Terminal-Bench (77.3%), token efficiency (4x fewer tokens than Claude Code), and kernel-level sandboxing. Most teams should start with Claude Code for complex projects, Gemini CLI for budget-conscious exploration, and Codex CLI for autonomous batch operations.
Last week, a client's CTO dropped a question in our Slack channel that I've been hearing from every engineering team this month: "We're standardizing on a terminal coding agent—which one?"
The terminal is the most contested real estate in developer tooling right now. Gemini CLI launched a free tier that undercuts everyone. Claude Code shipped Agent Teams that let multiple AI instances coordinate on the same codebase. Codex CLI posted a 77.3% Terminal-Bench score that neither competitor has matched. Six major comparison articles dropped in March 2026 alone, and most of them got the recommendation wrong because they optimized for benchmarks instead of workflows.
I've spent the last two weeks running all three on client projects—an Express.js API refactor, a Next.js migration, and a Python data pipeline rebuild. Here's what actually matters for picking one. For context on how these compare to IDE-based tools, see our Cursor vs Claude Code 2026 guide.
The Three Contenders at a Glance
Before diving into details, here's the landscape:
The benchmarks tell you these tools are converging on raw capability. The workflows tell you they're diverging on philosophy.
| Feature | Gemini CLI | Claude Code | Codex CLI |
|---|---|---|---|
| Developer | Anthropic | OpenAI | |
| Default Model | Auto-routes (Flash/3.1 Pro) | Opus 4.6 | GPT-5.4 |
| Context Window | 1M tokens | 200K (1M beta) | 192K tokens |
| Free Tier | 1,000 req/day | No | No |
| Entry Price | $20/mo | $20/mo | $20/mo (ChatGPT Plus) |
| SWE-bench Verified | 80.6% | 80.9% | ~80% |
| Terminal-Bench 2.0 | 68.5% | 65.4% | 77.3% |
| Open Source | Yes | No | Yes (Rust) |
| MCP Support | Yes | Yes | Yes |
| Sandboxing | bubblewrap + seccomp | Permission modes | Kernel-level (Seatbelt/Landlock) |
Gemini CLI: The Free Tier That Changes the Math
Gemini CLI's killer feature isn't a technical innovation—it's economics. One thousand requests per day, no credit card, no trial period. For a solo developer or a team evaluating terminal agents, this eliminates the cost barrier entirely.
Plan Mode Changes How You Start Work
Shipped with v0.34.0 in March 2026, Plan Mode is a read-only phase where Gemini CLI restricts itself to reading your codebase, asking clarifying questions, and proposing a strategy—without writing a single file. It sounds simple, but it addresses the most common failure mode of AI coding agents: jumping straight to implementation before understanding the problem. In practice, I've started using Plan Mode as a code review companion. Point Gemini CLI at a PR branch, enable Plan Mode, and ask it to identify risks. It reads every changed file, cross-references with the test suite, and flags potential issues—all without the temptation to "just fix it" before you've reviewed its reasoning.
PTY Shell Integration
This is Gemini CLI's most underrated technical feature. It spawns a virtual terminal (PTY) in the background, takes snapshots of terminal state, and renders output inline. This means it can run interactive tools—vim, htop, authentication prompts, install scripts that ask for confirmation mid-run. Neither Claude Code nor Codex CLI handle interactive terminal sessions natively. For our client running a database migration that required interactive confirmation prompts, Gemini CLI was the only tool that could execute the full migration script without manual intervention at the terminal level.
Where Gemini CLI Falls Short
The auto-routing between Flash and Gemini 3.1 Pro models is opaque. On complex refactoring tasks, I've seen it route to Flash when Pro would have been appropriate, producing shallow rewrites that missed edge cases. The 80.6% SWE-bench score comes from Gemini 3.1 Pro specifically—on the free tier's Flash model, expect significantly lower accuracy on complex tasks. Web search grounding is a double-edged sword. Gemini CLI can search the web mid-task to find documentation or examples, which is genuinely useful for unfamiliar libraries. But it occasionally hallucinates search results into code—citing a StackOverflow pattern that doesn't exist or referencing an API endpoint that was deprecated two versions ago.
Claude Code: When Accuracy and Coordination Matter Most
Claude Code's 80.9% SWE-bench Verified score is the highest of the three, but the number that matters more in practice is its reported 95% first-pass code accuracy. In our Express.js refactor benchmark, Claude Code finished in 1 hour 17 minutes with zero manual interventions—compared to 1 hour 41 minutes for Codex CLI and 2 hours 4 minutes with three corrections for Gemini CLI.
Agent Teams Are a Different Category
Launched with Opus 4.6 in February 2026, Agent Teams go beyond simple parallelization. Unlike subagents that report back to a single orchestrator, teammates communicate directly with each other through a shared task list and mailbox system. On a client's Next.js migration, we set up three teammates: one refactoring the API routes, one updating React components to match new data shapes, and one writing integration tests. The API agent discovered a type change that would break the frontend—and flagged it directly to the frontend agent, which adjusted its approach without us playing telephone. This kind of cross-agent coordination is something neither Gemini CLI nor Codex CLI offers. The tradeoff is token consumption. Agent Teams use roughly 4–7x more tokens than single-agent sessions. On the Max 5x plan ($100/month), a complex Agent Teams session can burn through your daily allocation in two hours.
Hooks and MCP Integration
Claude Code's hooks system lets you inject shell commands, HTTP calls, or LLM prompts at specific lifecycle points—when a subagent starts, when a file is modified, when a teammate goes idle. This bridges "let the AI figure it out" with "I need deterministic guarantees at specific steps." Combined with MCP integration for databases, Slack, GitHub, Sentry, and custom tooling, Claude Code becomes the most extensible option for teams with complex internal systems. We've configured Claude Code to automatically run type checks after every file edit and post a Slack notification when Agent Teams complete a task. Neither competitor matches this level of lifecycle control. For more on structuring AI coding agents with configuration files, see our guide on AGENTS.md configuration.
Where Claude Code Falls Short
The 65.4% Terminal-Bench score—lowest of the three—reveals a real weakness in raw terminal automation tasks. Claude Code excels at code understanding and generation but struggles with complex shell scripting, system administration, and terminal-based workflows compared to Codex CLI's 77.3%. No free tier means a $20/month commitment before you write a single line of code. And while the 1M token context is available with Opus 4.6, it's still in beta—the standard 200K window is what most users actually work with day-to-day.
Codex CLI: The Autonomous Terminal Specialist
Codex CLI wins Terminal-Bench 2.0 by a significant margin (77.3% vs. 68.5% and 65.4%) and does it while consuming roughly 4x fewer tokens than Claude Code for equivalent tasks. In a Figma-to-code benchmark, Codex CLI used 1.5 million tokens versus Claude Code's 6.2 million—producing comparable output at a fraction of the cost.
Full-Auto Mode with Kernel-Level Safety
Codex CLI's three approval modes—Suggest, Auto-edit, and Full-auto—are switchable mid-session via /mode. Full-auto removes all confirmation gates, letting the agent execute autonomously. What makes this viable rather than terrifying is OS kernel-level sandboxing: Seatbelt on macOS, Landlock plus seccomp on Linux. Network access is disabled by default in the sandbox. This means even if a prompt injection attack tries to exfiltrate code or hit an external API, the kernel blocks it. It's a fundamentally different security model than Claude Code's permission-based approach or Gemini CLI's bubblewrap isolation. For our Python data pipeline rebuild, we ran Codex CLI in full-auto mode for three hours straight. It refactored 47 files, ran the test suite after each change, and fixed its own test failures—all without a single human interaction. The kernel sandbox meant we didn't worry about it accidentally deleting the production database config.
Token Efficiency Matters at Scale
The 4x token efficiency gap is Codex CLI's most underappreciated advantage. For a team of ten developers each running 5–10 agent sessions per day, the difference between 1.5M and 6.2M tokens per session translates to thousands of dollars monthly. If you're on API pricing (GPT-5.4 at $1.25/$10.00 per 1M tokens vs. Opus 4.6 at $5.00/$25.00), Codex CLI is roughly 10x cheaper per equivalent task.
Open Source with Growing Ecosystem
Built in Rust with 67,000+ GitHub stars and 400+ contributors, Codex CLI is the most transparent of the three. You can audit the sandboxing implementation, contribute tools, and customize behavior at a level that Claude Code's closed-source architecture doesn't allow.
Where Codex CLI Falls Short
The 192K token context window is the smallest of the three—roughly 5x smaller than Gemini CLI's standard window. On large codebases, Codex CLI hits context limits faster, requiring more careful file scoping or chunked workflows. No Agent Teams equivalent. Codex CLI has subagents for task parallelization, but nothing matching Claude Code's direct agent-to-agent communication. For cross-cutting refactors that touch frontend, backend, and tests simultaneously, you're back to sequential orchestration. The pricing jump from $20/month (ChatGPT Plus) to $200/month (ChatGPT Pro) has no middle ground. Claude Code's $100/month Max 5x tier fills a gap that Codex CLI doesn't address.
Pricing Deep Dive: The Real Cost of Daily Use
The $20/month entry price is identical, but daily use economics diverge dramatically:
For teams evaluating tools, Gemini CLI's free tier is unbeatable. For sustained heavy use on API pricing, Codex CLI's token efficiency makes it 3–10x cheaper than Claude Code depending on the task.
| Usage Level | Gemini CLI | Claude Code | Codex CLI |
|---|---|---|---|
| Casual (5-10 req/day) | Free | $20/mo (Pro) | $20/mo (Plus) |
| Regular (50-100 req/day) | Free | $100/mo (Max 5x) | $20/mo (Plus) |
| Heavy (200+ req/day) | $50/mo (Ultra) | $200/mo (Max 20x) | $200/mo (Pro) |
| API (per 1M tokens) | $2-4 in / $12-18 out | $5 in / $25 out | $1.25 in / $10 out |
Decision Framework: When to Use Each
After two weeks of running all three across different project types, here's the framework I give our clients:
Choose Gemini CLI When:
- Budget is the primary constraint. The free tier handles most evaluation and learning workflows
- You need interactive terminal support. PTY shell integration handles prompts and interactive scripts that break other agents
- Large codebase exploration matters. The 1M standard context window means less chunking and file scoping
- Plan Mode fits your workflow. If you want AI to think before it writes, Plan Mode enforces that discipline
Choose Claude Code When:
- Code accuracy is non-negotiable. 80.9% SWE-bench and 95% first-pass accuracy mean fewer manual corrections
- Multi-agent coordination is needed. Agent Teams handle cross-cutting refactors that touch multiple system layers simultaneously
- You have complex internal tooling. MCP integration and hooks provide the deepest customization for enterprise environments
- Your team already uses [structured skill packs](/blog/superpowers-vs-gstack-ai-coding-skill-packs). Claude Code's skills and agents ecosystem is the most mature
Choose Codex CLI When:
- Autonomous execution is the goal. Full-auto mode with kernel-level sandboxing is the safest autonomous setup available
- Token cost matters at scale. 4x efficiency means 4x budget savings for large teams
- Terminal automation is the primary use case. 77.3% Terminal-Bench score means it handles shell scripts, system admin, and CLI workflows better than either competitor
- You want full transparency. Open-source Rust codebase with 400+ contributors means you can audit and customize everything
The Reality: Most Teams Will Use Two
The dirty secret of this comparison is that the tools complement each other more than they compete. Our own team runs Claude Code for complex client projects where accuracy and Agent Teams matter, Gemini CLI for quick explorations and planning sessions where the free tier keeps costs at zero, and Codex CLI for automated batch operations where token efficiency and sandboxing shine.
All three support MCP, so tool configurations are largely portable. The terminal agent category is converging on capabilities while diverging on philosophy—and that divergence is exactly what lets you pick the right tool for each task rather than forcing one tool to do everything.
The command line has become the most contested real estate in developer tooling. The good news is that every option is genuinely useful. The bad news is that you'll probably end up paying for two of them.
Frequently Asked Questions
Quick answers to common questions about this topic
Gemini CLI is the clear winner for free usage. It offers 1,000 requests per day with no credit card required, using the Flash model. Neither Claude Code nor Codex CLI offer a comparable free tier—both require $20/month subscriptions. For hobby projects, learning, or evaluation, Gemini CLI eliminates the cost barrier entirely.
