Codex vs Claude Code: Which One for an iPhone App?

Codex or Claude Code: the short answer

Codex or Claude Code: for an iPhone app built with Expo, both agents can read your project, edit files and run your commands. The difference that matters at the start is access: Codex has a limited free tier, while Claude Code requires at least a Claude Pro subscription. If you already pay for ChatGPT Plus or Claude Pro, go with the agent that's included. If not, try Codex for free, then decide using the criteria below.

Codex is OpenAI's coding agent, and Claude Code is Anthropic's. This comparison is based on their official pages, read on October 16, 2026, and on my workshop, where both agents have been working on my apps since summer 2026. You won't find any speed or quality ranking here: I didn't run a benchmark, and I'm not going to make one up.

Pricing, plans and usage limits

Prices are the ones shown in US dollars on the official pages. The amount you're billed in France depends on currency and taxes.

Criterion	Codex (OpenAI)	Claude Code (Anthropic)
Free access	Yes, Free plan: GPT-6 Luna in the desktop app, depending on rollout	No, Claude's free plan doesn't include Claude Code
Small plan	Go at $8 per month, same listed access as Free	None
Full entry plan	ChatGPT Plus at $20 per month	Claude Pro at $20 per month, or $17 per month billed annually ($200 paid upfront)
Heavy usage	Pro at $100, $200 or $500 per month	Max from $100 per month, with 5 or 20 times the usage of Pro
Team	Business at $20 per user per month billed annually, from 2 users	Team at $20 per seat per month billed annually, from 2 to 150 people
Pay as you go	API key at API rates, without the cloud features	Console account, at API rates
Limits	5-hour window and possible weekly limits; no 5-hour window on Pro for now	Rolling 5-hour window, plus weekly limits on paid plans
When you hit the limit	Buy credits on Plus and Pro	Usage credits at API rates on paid plans

Sources: the Codex pricing page, the Claude pricing page and the Claude Code requirements.

Three details really change your starting budget:

• Codex's free tier is narrow. For Free and Go, OpenAI's pricing page only lists GPT-6 Luna in the desktop app, "depending on rollout." The terminal, the editor extension, the web, iOS and GitHub code review show up starting with Plus. Image generation isn't available on Free.
• At Anthropic, everything draws from the same pool. Your conversations on claude.ai, on mobile and your work in Claude Code share the same limits. If you chat a lot with Claude during the day, you have less left for coding. At OpenAI, it's ChatGPT's Work mode that shares Codex's limits.
• Neither one promises a fixed number of messages. OpenAI gives estimates for Plus, for example 15 to 160 local messages per 5-hour window with GPT-6.1 Sol, and says these aren't fixed limits. Anthropic says usage depends on conversation length, the model and the features you use.

What they do in your project

Criterion	Codex	Claude Code
Interfaces	ChatGPT desktop app, terminal, VS Code, Cursor or Windsurf extension, web, iOS	Terminal, VS Code or Cursor extension, JetBrains, desktop app, web, monitoring in the mobile app
Operating systems	Desktop app on macOS, Windows and Linux; terminal on macOS, Linux and Windows	Terminal on macOS 13+, Windows 10+ and Linux; desktop app on macOS and Windows, Linux in beta
Models	GPT-6.1 Sol and GPT-6 Luna from Plus, GPT-6 Astra with fewer messages; GPT-6 Luna only on Free and Go	Opus 5.5 by default on Pro and Max, Sonnet 5.5 and Haiku 5.5 available, Fable 5.1 depending on plan
Running commands	Sandbox and approvals: "Ask for approval" mode works in your folder and stops before going outside it	Manual mode by default: permission asked the first time each tool is used; plan mode to explore without changing the code
Instructions file	AGENTS.md	CLAUDE.md, or AGENTS.md if the project has no CLAUDE.md
Screenshots	Image attached to the message, for example codex --image	Image pasted into the terminal with Ctrl+V
Local code review	/review: issues ranked by priority, without changing your files	/code-review: looks for bugs in your branch, on every plan
Pull request review	@codex review or automatic review on GitHub, from Plus	Automatic Code Review limited to Team and Enterprise; @claude with GitHub Actions

This table sums up the official Codex quickstart, its CLI, its editor extension and its permissions, along with the Claude Code overview, its models and its permissions.

What this means for you:

• The instructions file isn't a trap anymore. Codex reads AGENTS.md. Claude Code reads CLAUDE.md and, according to its documentation, it can read AGENTS.md when the project has no CLAUDE.md. So your project rules carry over from one agent to the other.
• Automatic pull request review is cheaper at OpenAI. Codex reviews your GitHub pull requests from Plus and, according to OpenAI, only flags P0 and P1 priority issues there. At Anthropic, the Code Review service requires a Team or Enterprise plan. On Pro, you're left with /code-review locally or a GitHub Action you set up.
• Model names change fast. OpenAI removed GPT-5.5 from Codex on October 14, 2026 for accounts signed in with ChatGPT, according to its models page. Don't pick an agent based on a model name: pick it based on access, limits and the way you work.

What matters for an iPhone app

This is where most comparisons stop too early. Building an iPhone app isn't just writing code: you have to see it run on a simulator or a phone, compile a build and get through Apple's review. On these steps, Codex and Claude Code have exactly the same limits.

Step	What the agent can do	What's still on you
Expo and React Native code	Read the project, edit screens, run npx expo start	Review the changed files
iOS Simulator	Launch it on your Mac with npx expo start --ios, if Xcode is installed	Have a Mac: the simulator only installs on macOS
Cloud or web session	Code on a copy of your repo, on a remote machine	Test it yourself: that machine doesn't run your local simulator
Swift project in Xcode	Work in Xcode 26.3, which supports Codex and Claude Agent	Choose the agent and accept its terms of use
EAS Build	Run eas build: the iOS build runs on Macs hosted by Expo	Approve each build, since it uses your Expo account
Testing on iPhone	Give you the list of things to check	Test on the real phone, including with the keyboard open and a small screen
App Store	Prepare the project and the submission commands	Apple Developer account, App Store Connect listing, Apple's review

Sources: the Expo documentation on the iOS Simulator, the EAS Build overview, Claude Code in the cloud and Apple's announcement about Xcode 26.3.

Neither agent holds your iPhone. A successful build doesn't prove a button stays visible when the keyboard is open.

For screenshots, both accept an image in the message. It's the fastest way to report a visual problem: you show the screen instead of describing it. Expo is also working on Expo Simulators, remote simulators an agent can drive, but it's still a limited-access preview, reserved for a few partners.

If you don't have a Mac, nothing stops you from using either one: you test on your iPhone and build with EAS. The guide publish an iOS app without a Mac walks through that setup.

My verdict for your situation

There's no universal winner. Here's what I'd tell a friend depending on their situation.

Your situation	My pick	Why
You want to try without paying anything	Codex Free	The only free entry of the two, limited to the desktop app
You already pay for ChatGPT Plus	Codex	Already included: terminal, editor, web and GitHub review
You already pay for Claude Pro	Claude Code	Already included, no second subscription
You follow my path or the iOS Kit	Claude Code	The rules and prompts are written for Claude
You want every pull request reviewed	Codex Plus	GitHub review included from $20; at Anthropic, you need Team or an Action
You code several hours a day	Compare Pro and Max	Both start at $100 per month; look mostly at the limits
You code in Swift in Xcode	The one you subscribe to	Both are built into Xcode 26.3

If you're a true beginner, the best agent is the one you'll stick with all the way. Switching tools every time you get stuck costs more than a subscription: you lose your rules, your habits and the thread of your project.

What I actually use in my workshop

I don't choose between them. Since summer 2026, Codex and Claude Code have been working in the same folder, on Solys, TradaZen and Nolera. Here's how it's organized, as I find it in my work files.

• One instructions file per agent. The folder has a CLAUDE.md for Claude Code and an AGENTS.md for Codex, with the same core rules.
• A handoff file. Before finishing, each agent writes down what it did and what's left. Strict rule: never two agents on the same files at the same time.
• One simulator each. Codex had its iPhone 17 Pro, Claude Code its iPhone 16e. They didn't steal each other's screen during screenshots.
• Separate roles. In August, Codex coded the Solys screens, and Claude Code handled the data, the logic and the checks. On Nolera, Claude Code took over the screens too.
• A relay when a limit hits. Nolera started with Codex while I'd hit a usage limit, then Claude Code took over the project.

The most useful lesson: no agent validates its own work. On TradaZen, Codex told me its screens were 9 out of 10 or higher. A separate review by Claude Code, on fresh screenshots from a small and a large iPhone, rated them between 6 and 8.5, with regressions. I don't draw a verdict on Codex from that, I draw a rule: since then, in my setup, the agent that codes a screen is never the one that grades it.

Right now, Claude Code does most of the coding, and the iOS Kit, its rules and its skills are written for it. I also use Codex to generate illustrations with my ChatGPT subscription.

The test that decides it for you

If you're still unsure, give both the same small exercise, on the same version of the project. For example, on a made-up shopping list app: add an item and keep it after the app is closed.

Make a Git checkpoint before each attempt, as OpenAI recommends, so you can roll back. Then note four things:

1. Did you understand the plan before it started coding?
2. How many files changed, and can you review them?
3. Does the agent separate what it checked from what you still need to test on the iPhone?
4. How much of your limit did the exercise use? Check /status in Codex and the Usage page in Claude's settings.

The install steps and first prompts are in the Codex tutorial and the Claude Code tutorial. I won't repeat them here.

How I built this comparison

• Facts checked on October 16, 2026 on OpenAI's and Anthropic's pricing pages and documentation, Apple's announcement about Xcode 26.3 and Expo's documentation. Every source is linked in the text.
• Real experience: my workshop, where both agents work on my iPhone apps. The examples come from my work notes.
• Limits: no speed or quality benchmark. Prices, models and limits change often. OpenAI's message estimates aren't fixed limits. Check your usage dashboard before committing to a plan.

Choosing what's next for your app

Also weighing Cursor or an app builder? The comparison of AI tools for building an app widens the choice.

If you go with Claude Code, the iOS Kit gives you the starter code from my apps and the prompts to give Claude at each step, all the way to the App Store. The Claude Code training page explains the method. With Codex, the vibe coding training page follows the same path, and the iOS app development training path works whatever agent you use.

Not sure the tool is your real blocker? The free diagnostic tells you what's actually holding you back, the idea, the code or Apple's review, then sends you my anti-rejection checklist.

Frequently asked questions

Is Codex free?

Partly. OpenAI's Free plan gives access to Codex with the GPT-6 Luna model in the ChatGPT desktop app, depending on rollout. The Go plan, at $8 per month, lists the same access. The terminal, the editor extension, the web and GitHub review are listed starting with Plus, at $20 per month. Image generation isn't available on Free.

Is Claude Code free?

No. Anthropic says Claude's free plan doesn't include Claude Code. You need a Pro, Max, Team, Enterprise or Console account. Claude Pro costs $20 per month, or $17 per month billed annually with $200 paid upfront.

Claude Code or Codex for beginners?

Pick the one you can pay for and stick with all the way. With no budget, Codex Free lets you try. If you already have ChatGPT Plus or Claude Pro, use the agent that's included. For an iPhone app, the difference then comes from your method: one feature at a time, reviewed changes and a test on your phone.

What is Claude Code's competitor?

For coding inside your own project, Claude Code's direct competitor is Codex, OpenAI's agent. Both read your files, edit them and run commands, in the terminal, an editor, a desktop app or the web. Cursor, a code editor with a built-in agent, is another option.

What is the best coding agent?

No general ranking applies to your project. For an iPhone app, the right agent is the one you can afford, whose changes you can follow and whose results you can check on your phone. Give both the same small exercise and compare what you understood, not the speed.

Why use Codex or Claude Code instead of a regular chat?

Because they work inside your project: they read files, edit them and run commands, with your approval depending on the mode you pick. In a regular chat, you copy every answer by hand. The tradeoff is that you have to review what they changed. The guide what is vibe coding explains this way of working.

Is Codex included in ChatGPT Plus?

Yes. According to OpenAI's pricing page, ChatGPT Plus at $20 per month includes Codex on the web, in the terminal, in the editor extension and on iOS, plus automatic code review on GitHub. Usage is limited per 5-hour window, and weekly limits may apply on top.

Can you use Codex and Claude Code on the same project?

Yes, that's what I do. Codex reads AGENTS.md, Claude Code reads CLAUDE.md and can also read AGENTS.md when the project has no CLAUDE.md. Give each one a role and never let both edit the same files at the same time. A handoff file, where each agent writes down what it did, prevents conflicts.

Do you need a Mac to use Codex or Claude Code on an iPhone app?

Not for coding: both agents run on macOS, Windows and Linux, and EAS Build compiles your iOS app on Macs hosted by Expo. The iOS Simulator, though, only installs on macOS. Without a Mac, you test on your iPhone.

Can Codex or Claude Code review my code?

Yes. In Codex, /review analyzes your changes and ranks the issues without changing your files. In Claude Code, /code-review looks for bugs in your branch, on every plan. On GitHub, Codex reviews your pull requests from Plus, while Anthropic's automatic review requires Team or Enterprise. An agent's review doesn't replace your test on the iPhone.

Can Codex or Claude Code publish my app on the App Store?

They can prepare the project and run EAS commands from your terminal. Publishing also goes through your Apple Developer account, the App Store Connect listing and Apple's review, which no agent can guarantee. The App Store publishing guide walks through these steps.