First-party only — a session id and where you arrived from, so we can tell which pages are worth writing more of. No advertising, no third-party trackers, nothing that follows you off this site.
Both are on this list because they are genuinely good at something. The question is which something you need. Below is the factual comparison, then the honest answer about who each one is for.
Reviewed July 25, 2026 · pricing changes often, check the vendor
| Codex CLI | Gemini CLI | |
|---|---|---|
| Made by | OpenAI | |
| Where it runs | Terminal | Terminal |
| Models | OpenAI models (GPT-5 family / codex variants) | Gemini models |
| Pricing | Included with ChatGPT Plus/Pro/Business · or API key | Free tier with a personal Google account · paid for higher limits |
| Open source | Yes | Yes |
| Best for | Teams already on ChatGPT plans who want a terminal agent with a clear sandbox story. | Trying a terminal agent seriously without a subscription, or working with very large contexts. |
The sandbox model is explicit and configurable, which makes autonomous runs less alarming
AGENTS.md is a plain convention other tools have adopted, so the config is portable
Local and cloud execution share the same interface
OpenAI models only
The cloud environment needs its own setup before it is useful on a private repo
Approval modes are powerful and easy to set too permissively
The free tier is real and large enough to do actual work
Very large context windows suit whole-repo questions
Apache-2.0 licensed, so you can read and fork it
Gemini models only
Younger than the alternatives; conventions are still moving
Free tier terms are subject to change
Pick Codex CLI if teams already on ChatGPT plans who want a terminal agent with a clear sandbox story.
Pick Gemini CLI if trying a terminal agent seriously without a subscription, or working with very large contexts.
And a thing worth saying plainly: the gap between these two is smaller than the gap between a repo that tells an agent what it is and one that does not. If you are choosing between them to fix inconsistent output, the tool is probably not the variable. The free Foundations track covers the part that is.