Using AI in your research
Claude Code and Codex, side by side
Claude Code, from Anthropic, and Codex, from OpenAI, are the same kind of tool: an agent that works inside a folder on your computer, reading and writing files and running commands until a task is done. Most of what you learn in one carries over to the other. The differences are in names, defaults, and plans.
Last verified November 4, 2026, against Claude Code 2.1.318 and Codex 0.166.0. Both change often. Typing / in either one lists its current commands.
Plans and data
| Claude Code | Codex |
|---|
| Plans that include it | Pro ($20 a month) and higher. Not the free plan. | Plus ($20 a month) and higher, and MSU’s ChatGPT Edu. Free and Go get it only in the ChatGPT desktop app, as it rolls out. |
|---|
| MSU data rules | Public data only. Claude isn’t approved at MSU yet, and MSU expects an agreement with Anthropic in December 2026. | On an Edu seat, up to confidential research, with MSU IT approval. On a personal plan, public data only. |
|---|
| Usage limits | A five-hour window and a weekly cap. /usage shows both. | A five-hour window, and weekly limits may apply. /status shows what’s left. |
|---|
More on data rules: Choosing a tool. Installing either one: Setting up the terminal.
Starting and stopping
| Claude Code | Codex |
|---|
| Start in this folder | claude | codex |
|---|
| Start with a task | claude "summarize the PDFs here" | codex "summarize the PDFs here" |
|---|
| Sign in | A browser opens the first time. /login switches accounts. | A browser opens the first time. Choose Sign in with ChatGPT. |
|---|
| Stop it mid-task | Esc | Esc |
|---|
| Leave | /exit | /exit |
|---|
| Pick up where you left off | claude --continue | codex resume --last |
|---|
| Choose an older session | /resume | /resume |
|---|
Talking to it
| Claude Code | Codex |
|---|
| Point it at a file | Type @ and start the file’s name. | Type @ and start the file’s name. |
|---|
| Run a command yourself | Start the line with ! | Start the line with ! |
|---|
| New line without sending | Shift+Enter | Shift+Enter |
|---|
| Search earlier prompts | Ctrl+R | Ctrl+R |
|---|
| Dictate instead of typing | Any voice-to-text tool works. See Setting up the terminal. |
|---|
Instructions and memory
| Claude Code | Codex |
|---|
| Instruction file for this folder | CLAUDE.md | AGENTS.md |
|---|
| Instructions for every project | ~/.claude/CLAUDE.md | ~/.codex/AGENTS.md |
|---|
| Draft a first instruction file | /init | /init |
|---|
| Bring over the other tool’s setup | /import | /import |
|---|
| Memory between sessions | On by default. It saves notes to reuse later. /memory shows them and turns it off. | Off by default. Turn it on in Codex’s settings, and /memories controls it for each chat. |
|---|
The instruction file is plain text you can read and edit. Whatever is in it shapes every session in that folder, so it’s the first place to look when an agent behaves differently than you expected.
Permissions and safety
| Claude Code | Codex |
|---|
| How it starts | Auto mode. A second model reviews each action instead of asking you. Anthropic says this does not guarantee safety. | In a Git folder, it edits files in that folder and asks before anything else. In other folders it starts read-only. Network access is off. |
|---|
| Change the mode | Shift+Tab cycles through Manual (asks first), accept edits, plan, and auto. | /permissions switches between Auto and Read Only. |
|---|
| Plan before acting | Plan mode, from Shift+Tab or /plan | /plan |
|---|
| Sandbox | Off by default. /sandbox turns it on (Mac, Linux, and WSL 2 on Windows). | On by default, on Mac, Windows, and Linux. |
|---|
| Undo changes | /rewind, or Esc twice on an empty prompt, rolls back files and the conversation. | Save Git snapshots before and after a task. /diff shows what changed. |
|---|
Everyday commands
| Claude Code | Codex |
|---|
| See all commands | /help | Type / |
|---|
| Check your usage | /usage | /status |
|---|
| Switch model | /model | /model |
|---|
| Set how hard it thinks | /effort, from low to max | In /model, from low to ultra |
|---|
| Free up room in a long chat | /compact | /compact |
|---|
| Start fresh | /clear | /new |
|---|
| See how full the context is | /context | /status |
|---|
| Review its changes | /review | /review |
|---|
| See connected tools | /mcp | /mcp |
|---|
Models
| Claude Code | Codex |
|---|
| Choices, November 2026 | Opus 5.5 (the default), Sonnet 5.5, Haiku 5.5, and Fable 5.1, the most capable, which costs extra on Pro. | GPT-6.1 Sol (OpenAI’s suggested start), GPT-6 Astra (the most capable), GPT-6.1 Terra (fast), and GPT-6 Luna (the most efficient). |
|---|
When you compare the two, use matched models and effort levels on each side.
Working at scale
| Claude Code | Codex |
|---|
| Run one task from a script | claude -p "..." starts in Manual mode, so it can’t take actions you haven’t allowed in advance. | codex exec "..." starts read-only and expects a Git folder. |
|---|
| Sub-agents | Ask for them, or define them as Markdown files in .claude/agents/ | Ask for them, or define them as TOML files in .codex/agents/ |
|---|
| Skills (saved instructions for a task) | .claude/skills/<name>/SKILL.md, run with /name | .agents/skills/<name>/SKILL.md, run with $name |
|---|
| Connect outside tools (MCP) | claude mcp add | codex mcp add |
|---|
| Web search | Built in. | On by default, from a cached index. Start with codex --search for live results. |
|---|
Other places it runs
| Claude Code | Codex |
|---|
| An app window, no terminal | The Code tab in the Claude desktop app (Mac, Windows) | Codex mode in the ChatGPT desktop app (Mac, Windows, Linux) |
|---|
| Inside VS Code | The Claude Code extension | The Codex extension |
|---|
| In the cloud | claude.ai/code, working on a GitHub repository | Codex Cloud, which MSU turns off for Edu seats |
|---|
