Codex vs Claude Code: two terminal agents compared

Containment, which drives everything else

Both put an agent in your terminal. The design disagreement is about how to stop it doing something you did not want.

The 2026 correction

The old summary of this comparison, "Codex sandboxes and Claude Code asks", is now a statement about defaults rather than capabilities. Claude Code has its own sandboxing for shell commands and an auto mode where a classifier reviews actions before they run, and Codex has an untrusted approval policy that will prompt on anything off a trusted list. Pick on default behaviour, because the default is what you will actually live with.

The models behind each one

This is where a lot of the felt difference comes from, and both lineups moved in 2026.

Codex making reasoning effort an explicit setting matters more than it sounds. It means the same model is three different tools depending on one config key, and it is the largest lever you have on both quota and wall-clock time. Claude Code puts the equivalent decision in model selection and in whether you use plan mode.

Cost

The ladders match at the $20 entry tier but differ above it, which makes plan shape part of the choice. ChatGPT has an $8 Go tier with no Claude equivalent, and publishes Pro at $200. Codex also sells credits for overage, so a heavy day can continue at a price, where a Claude subscription simply stops until the window refills.

How each one meters you

• Codex counts messages against a 5-hour window that scales with your plan. As of August 2026 OpenAI publishes an indicative range of roughly 25 to 200 local messages per window on Plus for the everyday model, and describes Pro as offering maximum Codex tasks. Cloud tasks share the same window.
• Claude Code counts tokens weighted by model against a rolling 5-hour window plus a separate weekly cap. There is no message count, which is why the same plan feels generous on a Haiku day and tight on an Opus one.

The two-subscription argument

At $20 plus $20 you get two independent quotas and two independent opinions. When you exhaust one, you switch rather than stop. For anyone doing this full time that is a better use of $40 than a single upgrade, and it is the single most common setup among heavy users.

What people consistently report

Filtering months of comparisons down to the claims that recur across independent sources rather than appearing once:

Running both, properly

Not "have both installed", but an actual division of labour. The patterns that hold up:

# Claude Code makes its own worktree
claude --worktree feature-auth

# Codex in a second worktree, with the main repo writable so git works
git worktree add ../myapp-codex -b feature-auth-codex
cd ../myapp-codex
codex --add-dir ~/code/myapp

Measuring which one is actually earning its keep

Neither vendor can tell you what the other is costing, so if you run both you have no total. Both write their history to disk in a form you can read, which is the only place a real answer exists.

du -sh ~/.codex/sessions/ ~/.claude/projects/
ls -t ~/.codex/sessions/ | head -5
ls -t ~/.claude/projects/ | head -5

That tells you volume, not dollars. Pricing it properly means joining each turn’s token counts to the rate for the model that produced them at the time it ran, which is what dedicated tooling exists to do.

Picking one

If you genuinely have no prior subscription and no preference, start with whichever ecosystem you already trust with your data. That constraint is more durable than any current quality gap, because the quality gap will invert at least twice before your subscription renews.

Questions people ask

Keep reading