Short answer: Gemini CLI to start free, Codex if you pay OpenAI, Claude Code if you pay Anthropic
Verified July 27, 2026. TL;DR: All three are agents that live in your terminal and edit your repo. The split is access, not ambition. Gemini CLI is the only one with a serious free allowance, Codex comes with the ChatGPT plan you probably already have, and Claude Code requires a paid Claude subscription.
| Gemini CLI | Claude Code | Codex CLI | |
|---|---|---|---|
| Free tier | Sign in with Google: 1,000 requests/day, 60/min | None | Included with ChatGPT Free |
| Entry paid tier | Gemini Code Assist Standard: 1,500 requests/day, 120/min | Claude Pro $20/mo, or $17 billed annually | ChatGPT Go $8/mo, Plus $20/mo |
| What it meters | Model requests per day and per minute | A rolling five-hour window plus a weekly window | A five-hour window, plus weekly limits |
| Models | Gemini | Claude | OpenAI |
| Pay as you go | Gemini API key or Vertex AI | Usage credits, or the Claude Console API | API key, billed at API pricing |
Only Gemini CLI is free at real volume
Sign in with a personal Google account and Gemini CLI gives you 1,000 model requests a day and 60 a minute at no cost. That is not a trial. Codex is included on ChatGPT Free, so it also costs nothing to start, though the higher limits sit on the paid tiers. Claude Code has no free tier and ships only with a paid Claude plan.
Requests versus hours: two ways to run out
Google counts requests. Your Gemini CLI ceiling is a number per minute and per day, so you always know what is left. Anthropic and OpenAI count time instead. Claude Code refills on a rolling five-hour window plus a weekly one, and Codex resets on a five-hour window with weekly limits on top. A request budget tells you how much work you have left; a window only tells you when it reopens.
Each one sells the same escape hatch
Hit the ceiling and all three point you at an API key. Gemini CLI moves to usage-based billing through a Gemini API key or Vertex AI. Claude Code offers usage credits so you can keep working past a plan limit. Codex lets you sign in with an API key and pay only for the tokens it uses. On all three, the subscription is the discount, not the cap.
You are picking a model vendor, not just a CLI
Each one defaults to its maker's models: Gemini CLI to Gemini, Claude Code to Claude, Codex to OpenAI. Codex is the one that bends, since its config file takes a custom provider and a --oss flag for a local Ollama or LM Studio model. The editors bend by design instead, and Cursor and GitHub Copilot sell several vendors' models as a headline feature.
How to choose, in one line each
Already paying for ChatGPT, install Codex. Already paying for Claude, install Claude Code. Paying for neither and curious what an agent does to your repo, start with Gemini CLI and spend nothing. If the real question is which editor to live in, read Claude Code vs Cursor or the three-way Codex, Claude Code, and Copilot breakdown.