Codex vs Claude Code: I Built One Feature With Both (2026 Field Log)

Quick answer (2026): OpenAI Codex and Claude Code are both terminal-first AI coding agents, and in 2026 they are close enough on raw model quality that the benchmark scores are the wrong thing to argue about. I built the same pagination change on one repo with each. Claude Code won when I wanted to stay in the loop and shape a multi-file change while it happened. Codex won when I could fully specify a chunk of work and delegate it to run on its own while I did something else. The real question is not which model scored higher this month. It is which harness fits the task in front of you.

The setup: one repo, one boring task

I wanted a fair test, so I picked a change that is real but not glamorous: add cursor-based pagination to a list endpoint in a small Next.js app, update the client hook that calls it, and backfill the tests. Three files that touch each other, plus a migration to add an index. The kind of task where the model is not the hard part. The coordination is.

I ran it twice on the same commit. First with Codex, then I reset and ran it again with Claude Code. Same prompt, same repo, same coffee.

What Codex did

OpenAI's Codex CLI (2026) runs interactively in the terminal, but the part that changed how I use it is codex cloud and codex exec. I can hand it a well-specified task and let it run asynchronously, either in a cloud environment or non-interactively in a script. Its default model in 2026 is a GPT-5.6-class Codex model with configurable reasoning effort.

So I wrote the task out properly. Endpoint contract, the cursor encoding I wanted, the test cases. Then I let it go and made lunch.

It came back with a complete diff. The pagination worked. The encoding was sane. It had even added a test for an empty cursor, which I had not asked for. Good.

The cost showed up at review time. I was reading a finished pull request I had not watched get built. One of the three files had a subtle bug: the client hook cached the first page and never invalidated on cursor change. Easy to miss in a diff, obvious the second I ran it. Because I had delegated the whole thing, I had no mental model of how it got there. I had to reconstruct the reasoning to fix five lines.

What Claude Code did

Claude Code (2026) is built around the interactive loop. It maps the codebase with agentic search, proposes the multi-file edit, and I steer as it goes. It runs in the terminal, in VS Code and JetBrains, and now on desktop and mobile too. Anthropic bundles it into the Claude Pro plan at 20 dollars per month and the Max plans at 100 and 200 dollars (pricing, 2026).

I stayed at the keyboard for this one. It read the three files, told me its plan, and I caught the caching problem before it was written, because it said out loud that it would memoize the first page. One sentence from me and that whole class of bug never happened.

The tradeoff is that I was there the entire time. Twelve minutes of my attention instead of zero. For a change this size that felt right. For a batch of ten of these, it would feel like babysitting.

The comparison, honestly

Both speak MCP (2026), so both plug into the same servers and neither locks you out of your tooling. Neither is a toy. Anyone telling you one is strictly better in 2026 is selling a benchmark, not describing a workday.

The reframe: it is a harness question, not a model question

The endless "has Codex overtaken Claude Code" threads treat this like a leaderboard. That is the wrong axis. The models trade the lead every few weeks and it barely changes my day.

What changes my day is autonomy style. Codex is happiest when I can fully specify work and delegate it, then review the result like a pull request from a fast contractor. Claude Code is happiest when the work is exploratory or entangled and I want to shape it as it forms. The delegation tax is real: the more you hand off unwatched, the more review you owe later, and review is where the subtle bugs hide.

What I kept

Both. That is the honest answer, and it is not a cop out.

Claude Code is my default for anything I am still figuring out, or any change that spans files I do not fully hold in my head. I reach for Codex when the task is boxed in tightly enough that I would rather spend the twelve minutes elsewhere, and I batch those. My routing rule now: if I can write the acceptance test before I start, it is a Codex job. If I cannot, I stay in Claude Code and steer.

If you only run one, pick the one that matches how you actually work, not the one that won last week's benchmark. And if you are still deciding your whole terminal-agent setup, my current lineup of the agents I reach for and the OpenCode versus Claude Code split cover the rest of the field I have tested this year.

Postscript: I still forget to reset the repo between runs about a third of the time. The agents are more disciplined than I am.