Codex CLI review: what beats Claude Code after 2 months

After two months of daily use, Codex CLI beats Claude Code on three things: token cost, GitHub PR reviews that catch real bugs, and headroom on the $20 plan. Claude Code still wins on long-horizon refactors. I run both and route work to whichever fits, so this is not a switch story.

I did not plan to keep both. The plan was a two-week trial of Codex CLI, then a verdict. The verdict turned into a split, because each tool is clearly better at a different part of the day.

Where Codex CLI beats Claude Code

Three places, in order of how much they changed my routine.

First, cost. Codex CLI burns fewer tokens for a comparable task, so the same work costs less per run. Second, PR reviews. Codex reviews a pull request on GitHub and flags bugs I would have shipped. Third, headroom. The $20 plan carries a heavy day for me, where I used to watch the meter on the equivalent Claude tier.

None of these is a benchmark number I am quoting from a vendor page. They are what I noticed using both tools on the same repos.

Token cost is the real difference

Codex CLI is the cheaper tool per task in my use. For pay-per-token work that is a smaller bill. For subscription work it means more runs before you hit a wall.

The mechanism is token volume, not the headline price. Codex tends to read less and write tighter diffs for the same change, so fewer tokens move per task. That compounds over a day of small edits. If you have already tried to cut Claude Code's token bill with model routing, this is the same lever from a different angle. The tool that reads fewer files wins on cost before you touch the model picker.

I saw a real gap. I did not measure a clean multiple across enough comparable tasks to put a number on it, so I am not going to invent one. The tidy multipliers you find quoted for this trace back to blog posts with no benchmark behind them.

The GitHub review bot earns its keep

The single feature that made me keep Codex is its code review on GitHub. It reviews a pull request and catches bugs, not style nits. OpenAI scopes it to P0 and P1 issues by design, which is why the comments stay on real risks.

In one PR it flagged an off-by-one in a pagination loop that my tests missed and I would have merged. That is the kind of catch that pays for the tool by itself. Claude Code's review integration was weaker for me here. It commented, but it commented on the obvious. Codex read the diff like someone who had been burned before.

This is the part I trust least to generalize. Review quality swings with the diff and the model version. But two months in, the Codex review is the one I read first.

Where Claude Code still wins

Claude Code keeps the work that needs a long memory.

Hooks used to lead this list. Codex shipped them to general availability in May 2026, enabled by default, so I no longer count hooks as a reason to reach for either tool.

Custom commands went the same way. Codex dropped its custom prompts in March 2026, and both tools now point you at skills you commit to the repo, so my stack of Claude Code commands is a habit rather than an edge.

What is left is long-horizon work. On a refactor that touches forty files across an afternoon, or an architecture change where one wrong assumption costs an hour, I reach for Claude Code. It holds the thread better. If the job needs delegation, how Claude Code handles sub-agents is still ahead of anything I get from Codex.

There is also the config split. Codex reads AGENTS.md. Claude Code reads CLAUDE.md. Run both on one repo and you maintain the CLAUDE.md vs AGENTS.md split, or you point one file at the other. It is a small tax, paid daily.

How I split work between the two

Codex CLI for cheap iterative edits, PR reviews, and anything where I want a second set of eyes on a diff. Claude Code for big refactors, scripted workflows, and the high-stakes changes where being wrong is expensive.

Routing by task type beat picking one. Two months in, I open Codex first for the small stuff and reach for Claude Code when the change gets serious. The cost section pays for both subscriptions, and the GitHub review alone justifies the second seat.