Codex vs Claude Code: a side by side comparison.

Prices in US dollars from claude.com/pricing and learn.chatgpt.com/docs/pricing on 23 October 2026, before tax. They change often, so check both pages before you buy.

On sticker price, Codex is cheaper to start: it is in every ChatGPT plan, including Free and the $8 Go plan, while Claude Code starts at Claude Pro. At $20 a month the two entry plans cost the same, and both top tiers start at $100. What the price does not tell you is how much work each plan carries before you hit a usage limit. Each vendor describes plan usage in its own terms (Claude Max as 5x or 20x Pro usage, ChatGPT plans by how many focused coding sessions they fit), so there is no like-for-like number to compare, and any "more usage per dollar" claim you read is a snapshot.

The cheapest agent is usually the one your existing subscription already covers. If you hit limits, a second plan on the other vendor costs the same as moving up a tier and gives you a second model family. When Codex says you have run out, Codex CLI usage limit reached explains the reset.

Neither is better across the board. Both vendors ship new models every few months, and a benchmark result or a "after 100 hours" verdict describes one model pair on one kind of task at one moment. We do not quote benchmark numbers here because they would be stale before this page is.

The differences that last are the ones in the table: the default sandbox and open source CLI on the Codex side, the accept edits and auto modes and checkpoint restore on the Claude Code side, and the plan each one rides on. The fastest way to settle it for your code is to give both the same task in the same repository and read the two diffs.

Yes: Codex CLI is OpenAI's equivalent of Claude Code, and most of what you do in one has a direct counterpart in the other. If you are moving across, this is the map. Claude Code can also pull your Codex setup in: claude import codex imports config from Codex, and --dry-run shows what it would write first.

Yes, and plenty of people do. They are separate programs with separate logins and separate session files, so both install on one machine and run at the same time. One AGENTS.md can serve both, since Claude Code reads it when there is no CLAUDE.md. A common split is a premium plan on the hard problems and a cheaper plan on routine work.

The fairest comparison is your own repository. Install both CLIs on a machine, connect it to Termdeck, and start two chats on the same folder: one on Claude Code, one on Codex, same prompt. Both appear in one sidebar, stream into the same transcript view, and stop for approval on the same cards. Each engine runs unmodified under your own login and keeps its own session files, so claude --resume and codex resume still work in a terminal afterwards. Use Claude Code and Codex in one UI walks through the setup step by step.

Do not point both at the same chat at once: two writers fork a transcript, so Termdeck keeps a session attached to another process view-only. Termdeck uses one set of permission mode names for both engines and prints the caveat where an engine differs.

No, but the name has been used twice. OpenAI first called a family of code completion models in its API Codex; those (code-davinci-002, code-cushman-001 and their siblings) were shut down on 23 March 2023, which is where most "Codex is deprecated" posts come from. Today's Codex is a different product: the coding agent in the CLI, the ChatGPT desktop app, the IDE extension and Codex Cloud. It is under active development and included in every ChatGPT plan.

Individual models that Codex runs do get retired on a schedule, as with any OpenAI model, so a model name pinned in your config.toml can stop working. OpenAI's deprecations page lists the dates. The same is true of Claude models in Claude Code.