Codex vs Claude Code: What I've Learned Using Both Daily

Resource footprint — what your machine actually feels

Before touching a single line of code, the two tools already feel different just sitting there. Claude Code's process is visibly heavier in Activity Monitor, routinely a few hundred MB, while Codex idles in the tens of MB. Typing claude at the terminal means a few seconds of waiting before the prompt is ready; typing codex is close to instant.

To put a number on that instead of just describing it, I measured both CLIs on the same machine (a MacBook Air, M3, 24GB RAM, macOS 26.5.1), same network. First, a trivial non-interactive task (a one-line reply prompt, no file edits) run back to back through both, using /usr/bin/time -l:

Then I launched each one interactively and just left it sitting at the prompt, untouched, for a few seconds before checking its resident memory with ps:

That idle number is closer to what you actually feel when both tools are just open in a terminal tab in the background — and it lines up with the gut-feel "hundreds of MB vs tens of MB" split almost exactly.

This isn't a rigorous benchmark, just one machine and one session — but the memory gap is consistent enough across both the task and idle numbers that it's not noise.

Pricing: what I'm actually paying for

Neither tool bills me per token unless I go through the API directly — both ride on their parent company's flat subscription tiers, and that's what decides how much room I have before I get throttled.

(Current as of Anthropic's pricing page and OpenAI's Codex pricing page — I'd check both before committing since tiers shift.) The gap I actually feel: OpenAI has no $100 middle tier, so once I'd be maxing out ChatGPT Plus, the next step is straight to $200; Anthropic gives me a $100 step in between. When I'm calling either model directly through the API instead of the CLI's bundled subscription — building my own agent rather than using the packaged tool — that's priced separately. For that use case I go through AIReiter rather than paying list price to Anthropic or OpenAI directly; it resells the same models at a discount.

Why I keep switching back and forth

The tool I reach for changes more with the underlying model than with the tool itself. When OpenAI's GPT-5.5 shipped, it was noticeably stronger than Claude for a stretch, and I moved most of my day-to-day work to Codex during that window. That advantage didn't last — GPT's quality dipped again a few weeks later, and I moved back to Claude Code as the default. Claude has its own quality swings too, but they've been smaller and less disruptive than what I've seen out of GPT over the same period.

Picking "Codex vs Claude Code" is partly a proxy for picking whichever frontier model is currently sharper that week. Evaluating only the CLI shell — the UI, the sandbox, the ecosystem — misses the variable that actually moves people between the two. It also maps onto the more commonly named split between the two tools: Codex leans hands-off (give it a task, walk away, review the diff later), while Claude Code leans hands-on (faster back-and-forth, more checkpoints where it waits on you).

Feature maturity gap — the missing rewind

Tool maturity is where Claude Code still has a clear lead over what I get from Codex, and the clearest example for me is one specific, unglamorous feature: rolling back to an earlier version of code mid-session. Claude Code supports this natively — I can rewind to a prior checkpoint when an edit goes sideways. Codex still doesn't, and it's been out long enough that I don't buy "just shipped, give it time" as an excuse anymore. Iterating fast and occasionally letting the agent go too far is exactly my workflow, so that's a feature I notice the absence of most weeks, not an edge case.

Claude Code's broader ecosystem carries the same pattern for me: subagents, hooks, and the Agent Skills system (a standard Anthropic introduced first, which Codex later adopted) give it more surface area to build workflows on top of. Codex's equivalent tooling exists but feels thinner in my day-to-day use, and its CLI/IDE/cloud/app surface is more fragmented in naming even as OpenAI has pushed it into more places.

Sandbox and the trust question

The two tools protect me from my own mistakes differently. Codex enforces isolation at the OS level — Seatbelt on macOS, Landlock on Linux, a Windows sandbox — so a runaway shell command is contained by the kernel, not by the agent's own judgment. Claude Code's safety model lives at the application layer, through a system of hooks I can configure to gate or block actions before they execute. Neither approach is strictly safer on paper; they're different failure modes, and Codex being open source means I can actually read how its sandbox is implemented rather than take it on faith.

That kernel-level isolation has a real cost in practice, though: Codex's sandbox pauses and waits on my explicit authorization for operations that Claude Code's hook-based model just lets through once I've configured it. Running a multi-step scripted workflow — the kind where a tool writes files or shells out to run a script partway through — means more interruptions asking for approval with Codex than with Claude Code.

Trust is also where Claude Code took a real hit this year, and it's the one thing that gives me pause about staying closed-source loyal. In July 2026, China's national vulnerability database publicly warned developers to drop recent Claude Code versions over a hidden mechanism in versions 2.1.91 through 2.1.196 that checked a device's timezone and whether traffic was routed through a non-official API endpoint, then flagged the request accordingly. Anthropic confirmed the mechanism existed, described it as an anti-distillation experiment rather than a backdoor, and removed it in version 2.1.198. I won't relitigate the whole incident here, but it's a real mark against a closed-source tool for me: I can't verify what I can't read, and Codex being open source means this specific failure mode doesn't apply to it the same way.

So which one should you actually use

If your day-to-day looks like mine — fast iteration on small-to-medium tasks, wanting a rich toolchain around you (subagents, hooks, skills, quick edits you can review and roll back as you go) — Claude Code is the safer daily default. If you're handing off a large, well-defined task you can walk away from, or you want an open-source agent you can audit and run inside a locked-down sandbox, Codex is a legitimate first choice, and it's the one I notice feels lighter on my machine while it works.

I don't fully commit to one, and neither do most heavy users I know. I default to whichever tool's underlying model is currently performing better, and keep the other installed for the jobs where its specific strengths — Codex's sandboxing and resource footprint, Claude Code's ecosystem and rewind support — actually matter for that task.

FAQ

Is Codex better than Claude Code?

Neither wins outright. Codex is lighter on resources and better suited to long, hands-off tasks; Claude Code has the more mature ecosystem, mid-session rewind, and a lower-cost middle subscription tier. Which one performs better on a given week often tracks the underlying model's current form more than the tool itself.

Is Claude Code or Codex free?

No — both require a paid plan for serious daily use. Claude Code needs Claude Pro ($20/mo) or higher; Codex needs ChatGPT Plus ($20/mo) or higher. Light, occasional use may fit inside free-tier trial limits, but neither sustains daily agentic coding for free.

Can I use both on the same project?

Yes, and it's a common pattern among heavy users: run one tool for the bulk of the work and have the other do a second pass — a fresh set of eyes catches blind spots the first tool's own model won't flag in its own output. Keep a short shared log of what changed and why so the second tool isn't guessing at context.

Bottom line

Both are capable daily tools with different weak points: Codex is lighter and faster to start but still can't rewind history mid-session; Claude Code is heavier on your machine but has the deeper toolchain, and its $100 mid-tier gives heavy users a step OpenAI doesn't offer. My own time split since both launched still lands above 75% Claude Code, and that's less a verdict on Codex than a reflection of how much the ecosystem gap matters for my actual workflow. Which one you should open today depends less on either tool's roadmap and more on which model behind it is currently doing its best work.