Ethan+Compare
Codex vs Claude Code: which one, and when to use both
I use Codex and Claude Code every day, often in the same build. Where each one is better, what limits actually decide, and how to use them together.
I use both of these every day. Most of my builds touch both. So this is not a feature checklist copied from two pricing pages. It is what the difference feels like when you are actually shipping, and why the answer for most serious users is not one or the other.
The short answer
- Pick Claude Code if you want the deepest control over how the agent works: instructions it always loads, reusable skills, hooks that enforce rules, subagents, and a large ecosystem of plugins and MCP servers. It shines in long sessions where you steer.
- Pick Codex if you live in the OpenAI ecosystem, already pay for ChatGPT, and want to hand off well defined tasks and get back a finished change. It is strong at delegation and fits naturally beside the rest of ChatGPT.
- Use both if you do long builds. Let one model plan and review and another implement. That is where the real savings are.
How they compare
| Claude Code | Codex | |
|---|---|---|
| Made by | Anthropic | OpenAI |
| Models | Claude models | OpenAI models |
| Where it runs | Terminal, IDE extensions, desktop and web | Terminal, IDE extension, desktop app and cloud tasks |
| Project instructions | CLAUDE.md |
AGENTS.md |
| Skills | Yes | Yes |
| Subagents | Yes | Yes |
| Hooks | Yes, a mature system | More limited |
| MCP tools | Yes | Yes |
| Paid through | Claude plans or the API | ChatGPT plans or the API |
| Best at | Long, steered sessions with a lot of customization | Delegated, well scoped tasks |
Both move fast, and features land in one and then the other within months. Treat the table as the shape of each tool, and check each tool's documentation for the details.
Where Claude Code is better
Extensibility. Claude Code's model of instructions, skills, hooks, subagents and plugins is the most complete I have used. You can make it work the way your team works and keep it that way. I wrote about getting the most out of that in How to build Claude Code skills that actually work.
Steering a long session. When I am exploring a problem, changing direction and making judgment calls as I go, Claude Code is where I want to be.
Following a process. Claude models handle written, multi-step processes well. Orchestration skills I wrote for other agents ran smoothly in Claude and were awkward elsewhere.
Where Codex is better
Delegation. Give Codex a clear brief and it goes and does it. Native subagents can be pinned to a specific model, which makes it a good host for a planner plus worker setup.
Living inside ChatGPT. If your company already runs on ChatGPT, Codex sits next to everything else you use, and there are ways to get more agent work out of the same plan. I wrote up one in How I run a coding agent on GPT-6 Pro inside ChatGPT.
Bringing other models in. With a router, Codex can send work to models from other providers as native subagents, while your main conversation stays on the model you trust for judgment.
What actually decides it: limits
For most heavy users the deciding factor is not features. It is how fast you hit your usage limit.
The expensive move is leaving your best model in the loop supervising: dispatching, waiting, re-reading, re-reviewing. In one long build I kept the expensive model on planning and final review only, gave implementation to a much cheaper worker, and cut my usage of the expensive model by roughly 95% without losing quality. A seven hour run used about 2% of my weekly limit. A five hour run done the old way had used over 28%. The full write-up is in How I cut my Astra usage.
That is a workflow change, not a tool change, and it works in either tool.
How I use both
I built Codex on Crack, an open source package, so one Codex conversation can plan, delegate to other models, and review their work. One of the arrangements it supports has Claude implement a whole assignment through Claude Code while Codex coordinates and checks the result with its own tools. My open source sales agent, Harvey, runs on either a Claude or a Codex subscription.
The tools are converging. The skill that lasts is knowing how to direct agents, split work between models, and review what comes back. That is what I teach teams in Claude Code and Codex training.
Questions people ask
- Is Codex better than Claude Code?
- Neither is better at everything. Claude Code has the deeper extension model for long, interactive sessions you steer yourself. Codex is strong for delegating well defined work and fits naturally if you already pay for ChatGPT. The model underneath and your usage limits often matter more than the tool.
- Can I use Codex and Claude Code together?
- Yes, and for long builds it is often the best setup. One model plans and reviews, another implements. I built Codex on Crack, an open source package, to coordinate models like this from one Codex conversation, including handing an assignment to Claude Code.
- Do Claude Code skills work in Codex?
- Both tools support skills as folders of instructions and resources, and the format is close enough that many skills move between them with little change. They are discovered from different locations, so check each tool's documentation for where to put them.
- Which is cheaper, Codex or Claude Code?
- Both are included in paid plans from their makers, starting at the entry tier, and heavy use pushes you to higher tiers. What it costs you depends on your limits and how you use them. Keeping the expensive model on judgment and delegating implementation cuts usage more than switching tools does. Check current prices on OpenAI's and Anthropic's pricing pages.