Codex vs Claude Code: which coding agent to pick
Published
Both are terminal coding agents that read your project, edit files, run commands and check their own work, and both start at US$20 a month (about CA$27.92). Pick Codex if you already pay for ChatGPT, want an open-source CLI, or want to hand tasks to a cloud sandbox. Pick Claude Code if you will spend most of the day in it and want the deepest set of controls for shaping how it works — instructions, memory, skills, hooks and subagents. The tool matters less than the setup you give it.
At a glance
| Aspect | Claude Code | Codex |
|---|---|---|
| Cheapest plan with it | Claude Pro, US$20 / month | ChatGPT Free (limited); Plus US$20 / month |
| Instructions file | CLAUDE.md (reads AGENTS.md too) | AGENTS.md, 32 KiB combined by default |
| Where it runs | terminal, IDE, desktop app, web | terminal, IDE extension, cloud tasks |
| Source code | closed | CLI is open source (Apache-2.0) |
What each costs in Canada
Codex is not sold on its own: it comes with a ChatGPT plan, the same way Claude Code comes with a Claude plan. Both vendors bill in US dollars.
| Plan | USD → ≈ CAD |
|---|---|
| ChatGPT FreeCodex included with lighter models and lower limits | US$0 |
| ChatGPT GoCodex included | US$8≈ CA$11.17/ month |
| ChatGPT PlusCodex included; usage counted in five-hour windows, weekly limits may apply | US$20≈ CA$27.92/ month |
| ChatGPT Protiers at US$100, US$200 and US$500; no five-hour limit | from US$100≈ CA$139.60/ month |
| ChatGPT BusinessUS$20 billed annually; two users or more | US$25≈ CA$34.90/ user / month |
| Plan | USD → ≈ CAD |
|---|---|
| Pro (monthly)the entry point for Claude Code | US$20≈ CA$27.92/ month |
| Max5x Pro's usage, every five hours and every week | US$100≈ CA$139.60/ month |
| Ultra10x Pro's weekly usage; no five-hour limit on verified accounts | US$200≈ CA$279.20/ month |
| Team, standard seatUS$20 / seat / month billed annually; teams of 2–150 | US$25≈ CA$34.90/ seat / month |
| Team, premium seatUS$100 / seat / month billed annually; 5x a standard seat's usage | US$125≈ CA$174.50/ seat / month |
At the entry tier the two are priced the same (US$20 against US$20). The difference is in how the allowance is counted. OpenAI publishes Codex limits per model, in messages per five-hour window, and cloud tasks draw more of the allowance than local work. Claude counts usage across chat and Claude Code together, in a five-hour window with a weekly cap. Either way, heavy daily use is what moves people to the higher tiers.
Where they actually differ
We run Claude Code every day across more than a dozen projects, and we have used Codex far less — so on the Codex side we stick to what its documentation says, and on the Claude side we write from use.
- Project instructions. Codex reads
AGENTS.mdfiles from your global folder, the repository root and the folders down to where you are working, up to a size limit. Claude Code readsCLAUDE.mdand can import other files into it, which is how we keep a short core with longer reference material loaded on demand. Claude Code can also read an existingAGENTS.md, so one file can serve both tools. - Controls around the model. Both now document skills, subagents, hooks and permission settings. In Claude Code these are the parts we lean on hardest: hooks that block a risky command before it runs, skills for checklists we would otherwise repeat, subagents that search or review without filling the main session’s context, and plan mode for anything large, where Claude proposes and we approve before it builds.
- Cloud work. Codex treats the cloud as a first-class place to send a task and come back later. Claude Code has the same idea on the web and in the desktop app; we still do most work locally, where it can reach our own tools and data.
- Openness. The Codex CLI is open source. If you want to read or patch the agent you run, that settles it.
The habit that matters more than the tool: write down what the agent learned before the session ends. Neither tool remembers a conversation by itself; both remember files. Our sessions append to a daily log as they work, and a long session that ends with nothing written is the failure we guard against most.
Who should pick which
- You already pay for ChatGPT: start with Codex. It is already in your plan, and trying it costs nothing extra.
- You want to delegate tasks and review pull requests later: Codex’s cloud tasks fit that rhythm well.
- You will work alongside the agent most of the day: Claude Code. Its instructions, skills, hooks and subagents reward the time you put into shaping them.
- A team with mixed preferences: keep project instructions in
AGENTS.mdso both tools read the same rules, and let people choose.
What we checked
Read 2026-11-01 from OpenAI's Codex pricing, the AGENTS.md guide, the Codex repository, claude.com/pricing and the Claude Code docs. CAD = USD × 1.396, the Bank of Canada daily rate for 2026-10-30 (Valet API), rounded to the cent. Before tax and card fees.
Getting it set up
Whichever you choose, the first afternoon decides whether it sticks. Our free guide to setting up Claude Code properly covers that afternoon, and most of it applies to Codex as well. AI Setup is that afternoon done with you, on your own work.