Terminal Wars: Claude Code vs Codex vs Gemini CLI in 2026

A year ago, AI coding meant autocomplete in your IDE. Tab, accept, move on. That era is over. The terminal is now the primary battlefield for AI coding agents, and three companies are fighting for it with fundamentally different philosophies.

Claude Code from Anthropic. Codex from OpenAI. Gemini CLI from Google. All three run in your terminal. All three can read your codebase, write files, and execute commands. But the similarities end there.

How They Actually Work

Claude Code runs interactively. You type a request, it thinks, shows you its reasoning, asks follow-up questions when it’s uncertain, and makes changes you can approve or reject in real time. It feels like pair programming with someone who reads fast. The model powering it (Opus 4.6 or Sonnet 4.6) runs locally in your terminal session, maintaining context across the conversation.

Codex takes the opposite approach. You give it a task and it disappears into a sandboxed environment, works autonomously for up to 30 minutes, then comes back with results for you to review. It’s closer to delegating than collaborating. The Codex macOS app, launched February 2, 2026, leans into this further: you can manage multiple agents working in parallel, each on its own worktree of the same repo.

Gemini CLI splits the difference. It’s interactive like Claude Code but optimized for speed over depth. Google positioned it as the fast-iteration tool: get to a working prototype quickly, then refine. It launched in June 2025 and accumulated over 85,000 GitHub stars within weeks.

The Feature Race

All three tools have been shipping at a breakneck pace. Here’s where each stands as of late February 2026:

Feature	Claude Code	Codex	Gemini CLI
Multi-agent	Agent Teams (research preview)	Multi-agent via macOS app	Not yet
Context window	200K (1M in beta)	200K	1M
MCP support	Yes	No	Yes
IDE integration	VS Code, JetBrains	VS Code extension	VS Code (via Gemini Code Assist)
Background agents	Yes	Yes (default mode)	No
Open source	No	CLI is Apache 2.0	Apache 2.0
Web search	No (via MCP)	Built-in	Built-in (Google Search)
Image input	No	Yes (screenshots, wireframes)	Yes

Claude Code’s recent additions are worth noting individually. Agent Teams, launched alongside Opus 4.6 on February 5, lets a lead agent delegate to multiple sub-agents working in parallel worktrees. The LSP tool, added in December, gives it language-aware code navigation in 11 languages, bringing search times down from 45 seconds to 50 milliseconds. And Claude Code Security, announced February 20, uses Opus 4.6 to hunt for vulnerabilities in codebases, a feature significant enough to crash cybersecurity stocks.

Codex’s biggest move was the Rust rewrite of its CLI. The original TypeScript/React version required Node v22+. The Rust rewrite eliminated that dependency entirely: zero-dependency install, lower memory usage, no garbage collection pauses. The macOS app adds scheduled automations, letting you set tasks to run on a timer with results queued for review.

Gemini CLI’s trump card is extensions (GEMS), which package prompts, MCP servers, custom commands, and agent skills into shareable bundles. Launch partners include Dynatrace, Elastic, Figma, Harness, and Stripe. The FastMCP integration since v2.12.3 makes building custom MCP servers significantly easier.

The Model Gap

The underlying models matter as much as the tooling.

Claude Code defaults to Sonnet 4.6, released February 17, with Opus 4.6 available for heavier tasks. Sonnet 4.6 introduced adaptive thinking, where the model dynamically decides how much reasoning to apply. Developers with early access preferred it to its predecessor “by a wide margin,” according to Anthropic.

Codex runs on GPT-5.3-Codex by default, released February 5. It scores 56.8% on SWE-Bench Pro and is 25% faster than its predecessor GPT-5.2-Codex. OpenAI describes it as the first model that was “instrumental in creating itself.” It’s also the first OpenAI model rated “High” in cybersecurity risk under their Preparedness Framework.

Gemini CLI uses Gemini 2.5 Pro (with Gemini 3 Flash recently added for faster workflows). Gemini 2.5 Pro ranked #1 on WebDev Arena for building web apps, and scores 63.8% on SWE-Bench Verified with a custom agent setup. The 1 million token context window comes standard, not as a beta feature.

Pricing: Google’s Free Tier Changes the Calculus

This is where the competition gets interesting.

Gemini CLI is free. Sixty requests per minute, 1,000 per day, with a personal Google account. Access to Gemini 2.5 Pro with a million-token context window. Google calls this “the highest free allowance available for AI developer tools,” and they’re not wrong.

Claude Code requires a paid plan. The Pro tier at $20/month gives limited access. Max 5x at $100/month and Max 20x at $200/month provide full access with increasing priority. API users pay per token: $3/$15 per million for Sonnet 4.6, $5/$25 for Opus 4.6.

Codex also has a free tier, at least for now. OpenAI is offering Codex for free via ChatGPT Free and Go plans for a limited time. Paid plans start at $20/month (Plus, with 45-225 local messages per 5 hours) and go up to $200/month (Pro, with 300-1,500 local messages per 5 hours). The CLI itself is open source under Apache 2.0, but you need either a ChatGPT subscription or an API key to use it.

So two of the three major terminal agents now have free tiers, and Claude Code is the outlier requiring a paid plan. For hobbyists and learners, that’s a real consideration. For professional developers already paying for Claude or ChatGPT, the marginal cost of the coding agent is zero either way. The pricing battle really matters at the enterprise tier, where Anthropic’s $2.5 billion annualized Claude Code revenue suggests they’re winning the contracts that matter most.

What Developers Actually Say

A UC San Diego and Cornell survey from January 2026 found that among 99 professional developers, Claude Code (58 users), GitHub Copilot (53), and Cursor (51) were the most widely adopted. Gemini CLI and Codex didn’t crack the top three, though both were growing.

The usage patterns differ. Claude Code users tend to run long sessions on complex tasks: multi-file refactoring, architectural analysis, debugging across codebases. Codex users lean toward the “fire and forget” model, assigning discrete tasks and reviewing results. Gemini CLI users skew toward rapid prototyping and MVP generation.

Twenty-nine respondents used multiple agents simultaneously. The hybrid approach, using Claude Code for feature generation and Codex for code review, keeps showing up in developer discussions as a common workflow.

The Third-Party Controversy

On January 9, 2026, Anthropic silently blocked subscription OAuth tokens in third-party tools like OpenCode, Roo Code, and Cline. Developers on $200 Max subscriptions who had been routing their usage through alternative interfaces were cut off without warning.

The backlash was immediate. OpenCode, with over 107,000 GitHub stars, was suddenly unusable with Claude. Developer sentiment was “overwhelmingly negative” according to coverage, primarily because Anthropic gave no advance notice.

OpenAI’s Thibault Sottiaux publicly endorsed using Codex subscriptions in third-party tools, drawing a direct contrast. Google’s Gemini CLI, being open source with a free API tier, sidesteps the issue entirely.

This is a real differentiator for developers who want flexibility in their tooling. Anthropic’s position is that third-party tools were generating disproportionate compute costs through Max subscriptions. The economics make sense from their side. But the way they handled it damaged trust with a vocal segment of their user base.

Which One Should You Use?

There’s no universal answer, but the use cases are becoming clearer:

Choose Claude Code if you work on complex codebases and want an interactive collaborator. It’s the strongest at multi-file refactoring, architectural reasoning, and long-running sessions. The Agent Teams feature is still in research preview but already useful for parallel task execution. If you’re an enterprise team, Anthropic’s 29% market share in enterprise AI and the Microsoft deployment (engineers across Windows, Teams, and Outlook are testing Claude Code alongside Copilot) suggest the enterprise tooling will keep improving.

Choose Codex if you prefer delegation over collaboration. Set a task, go work on something else, come back to review. The macOS app’s multi-agent management is the most polished of the three, and the Rust CLI is genuinely fast. GPT-5.3-Codex’s benchmark numbers are competitive, and the Skills system makes it easy to configure team-specific workflows.

Choose Gemini CLI if you want the lowest barrier to entry or need the largest context window without paying. The free tier with 1M tokens of context is unmatched. It’s also the best choice if you’re already deep in Google’s ecosystem. The GEMS extension system shows maturity, and Google’s investment in Jules (their autonomous coding agent) suggests the CLI will keep getting more capable.

Or use all three. That’s what an increasing number of developers are doing.

Sources: