Codex vs Claude Code: the verdict from running both

Short answer: Claude Code. On technical ability — long multi-file work, context handling, and how far it gets before it needs you — it is the stronger runtime, and it is what most people run on Fredrin. Pick Codex if your team already lives in ChatGPT, or if you want agents supervised from a GUI rather than a terminal.

Below is the long version of Codex vs Claude Code, including the places Codex wins outright.

Why can we answer Codex vs Claude Code straight?

Because Fredrin is not in this race. Both are runtimes our board launches, next to Cursor's agent and Copilot, and bring-your-own-subscription means you pay OpenAI or Anthropic directly either way. We see both execute on real repositories every day and earn exactly the same whichever one you choose.

OpenAI cannot publish this comparison. Neither can Anthropic. Most of the pages that try were written by people who have not run either tool for a full week on a codebase they have to maintain. We have no incentive to pick a winner and no reason to hedge to a tie, so the rest of this is what we see.

What is actually different between them?

Less than the marketing implies. Both are terminal-first agents with an IDE extension, a desktop app, a cloud mode and a project instructions file. Both read your repository, run your commands and open pull requests. The differences that matter are how much context each holds, how far it gets unsupervised, and where the review happens.

The surface argument is over. Claude Code runs in the terminal, an IDE, a desktop app and a browser; the Codex CLI covers the same ground from the other side, with codex exec for pipelines and /init writing you an AGENTS.md. Anyone still framing this as "terminal tool versus cloud product" is quoting last year.

Where is Claude Code stronger?

On long-horizon work. Hand it an outcome — make the suite green, port this module — and it keeps a coherent plan across dozens of files and tool calls without drifting or asking you to restate the goal. That endurance, not raw code quality on a single function, is the gap we see most consistently.

Three specifics behind that, in the order they show up:

• It holds a bigger working picture. On a task spanning fifteen files, Codex is more likely to lose the thread around file nine and start re-deriving decisions it already made. Claude Code carries them. You notice it as fewer "actually, undo that" turns.
• Its project configuration is deeper. CLAUDE.md, hooks that fire shell commands around its actions, skills that package a workflow, subagents that keep exploration out of the main context — all of it lives in the repo, so your teammates inherit it. Codex has skills, plugins and AGENTS.md; it does not yet have an equivalent of hooks.
• It recovers better from its own mistakes. Both write failing code. Claude Code is more likely to read the failure, form a hypothesis and fix the actual cause rather than patching the symptom and moving on.

None of that makes it magic. An agent that works unattended across a dozen files hands you a diff you have to genuinely read, and reviewing machine-written diffs without rubber-stamping them is the skill that decides whether any of this is faster.

Where is Codex genuinely better?

Supervising several agents from a screen instead of a terminal. OpenAI folded Codex into the ChatGPT desktop app and calls it "your command center for complex work", with inline diffs and a queue you scan rather than a stack of terminal panes. If you do not want to live in tmux, that is a real advantage.

Two more, honestly:

• The review loop is closer to the tool. /review is a first-class command in the Codex CLI, and code review in the cloud is listed as part of the plan rather than a separate integration you wire up.
• It is already paid for. If your company bought ChatGPT seats, Codex arrives with them. That is not a technical argument, and it wins a lot of decisions anyway.

On small, well-specified edits — rename this, add a guard clause, write this test — the two are close enough that habit decides it. The gap only opens as the task gets longer.

Codex vs Claude Code pricing: what does each cost?

Both are subscriptions, and the two price sheets have converged almost exactly. Codex is bundled into ChatGPT: OpenAI's docs list it as included in the Free, Go, Plus, Pro, Business, Edu and Enterprise plans, with Plus at $20 a month. Claude Code is bundled into Claude the same way, at $20 on Pro.

The top individual tier is the same shape too. OpenAI sells Codex Pro "From $100 /month" for "5x or 20x more Codex usage than Plus"; Anthropic sells Max "From $100" for "5x or 20x more usage than Pro". Both also offer a metered escape hatch — ChatGPT credits and an API key on one side, API billing on the other — for the automation that would otherwise eat a seat.

Two caveats. These pages move, so check them rather than trusting a post that was accurate the day it shipped. And neither company publishes its limits in units you can plan a sprint against, so a precise tasks-per-dollar figure is a guess. We ran the same arithmetic on Claude's own tiers; at one terminal the plans land in the same neighbourhood.

What do people actually run?

Claude Code, by a clear margin. We are deliberately not publishing a percentage: the honest version of this is an observation from watching the board, not a measured statistic we have signed off for publication. Runtime choice is configuration, not conversation content, but a number we have not verified is still a number we should not print.

The shape of it has been stable for months, though. New projects arrive on whichever tool the developer already pays for; the ones that end up running several tickets a day drift toward Claude Code, and the reason given is almost always endurance rather than output quality on any single change.

What does neither one fix?

The part after the agent stops. Both can now run several isolated sessions at once, so spawning is solved. What neither carries is a unit of work that survives the session: a branch, a diff, CI, a review decision, a merge. Finish five tasks in parallel and you have five sessions and no queue.

Three things fall on you in both tools:

1. Environment per checkout. A fresh worktree has no .env, no node_modules, no database and no free port. Claude Code documents --worktree and what a new tree does and does not inherit; the setup script that makes one usable is still yours to write. We wrote up why a fresh worktree starts life broken separately.
2. Unordered review. Six finished agents means six diffs and no opinion about which to open first. A session list sorts by recency, which is the one ordering unrelated to what needs you.
3. State in your head. Which branch was that on, did CI pass, did I already send it back and what did I say. A session is a conversation, not a unit of work, so there is nothing to hang that on.

The ChatGPT desktop app is the closest either vendor has come to this, and it is a good answer to watching parallel work — but it is still not a place a ticket lives from plan through merge. That gap is transparently what Fredrin exists to fill: one ticket, one branch, one worktree, one agent session, and a board that knows running from waiting-on-you from merged.

So which should you pick?

Claude Code if you delegate whole outcomes and read the diff afterwards; Codex if your team already pays for ChatGPT and wants agents supervised in a window rather than a terminal. Both are $20 to try for a month, they do not conflict on the same repository, and a week of real work settles it faster than any comparison.

• Long refactors, migrations, unattended multi-file work → Claude Code.
• A team already standardised on ChatGPT → Codex, and the saved procurement round is worth more than the delta.
• Small edits all day, one at a time → either; pick the one whose keybindings you already know.
• Three or more tasks running at once → either for the running, something with a board for the shipping.

If you land on Claude Code, the settings that decide how far it gets unattended are where the return actually comes from — and if you are weighing it against an editor rather than another agent, that is how Claude Code lines up against Cursor.

FAQ

Is Claude Code better than Codex? For long, multi-file work you hand over and check afterwards, yes — it holds a plan across more of the task before it needs you. For short edits the two are close, and for supervising several agents from a GUI the ChatGPT app is ahead.

Codex CLI vs Claude Code — how different is the terminal? Less than you would expect. Both are a REPL in your repository with slash commands for model, permissions and review, both take a project instructions file, and both have a non-interactive mode for CI. Codex uses AGENTS.md and codex exec; Claude Code uses CLAUDE.md and claude -p.

Codex or Claude Code — which is cheaper? At the entry tier they are the same: $20 a month, each bundled into the vendor's consumer subscription rather than billed per token. The top individual tier is also the same shape, from $100 a month for 5x or 20x the usage. Price is not the deciding variable here.

Can I run both on the same repository? Yes. They are separate CLIs reading separate instruction files, and neither locks the working copy. Keeping CLAUDE.md and AGENTS.md in sync is the only real cost, and Claude Code will read an AGENTS.md if that is what your repo already has.

Should I switch from Codex to Claude Code? Only if you are hitting the ceiling this post describes — tasks that stall halfway, or too many "you already did that" corrections. If your work is short, well-scoped edits and Codex is already paid for, switching buys you very little.

Fredrin is a desktop kanban for running many coding agents at once — a branch, a worktree and an agent session per ticket, then review and ship. It runs Claude Code, Codex, and a few others on whichever subscription you already pay for: fredrin.com.