Cursor vs Claude Code vs Codex: coding agents compared
Read in : Français

Contents
- The short answer
- The comparison table
- What does each tier cost?
- The differences that actually matter
- Interface: seeing the code or talking to it
- Safety defaults
- Project memory
- Background work
- The decision tree
- Same task, three ways to run it
- Frequently asked questions
- Frequently asked questions
- Sources
Cursor vs Claude Code vs Codex: all three coding agents can read a project, edit several files and run commands. The real differences are the interface (editor or terminal), which models the price includes, and how you hand tasks off to the cloud. This comparison is based on the official documentation and pricing pages, checked on November 21, 2026.
The short answer
Pick Cursor if you want a complete visual editor, Claude Code if you want Claude models for a flat subscription, and Codex if you already pay for ChatGPT.
- Cursor: a code editor built on VS Code, with autocomplete, an agent, Plan Mode and cloud agents. The most comfortable if you want to see and touch the code.
- Claude Code: Anthropic's agent, in the terminal, VS Code, JetBrains, the desktop app and the web. Strong on long tasks and fine-grained setup (CLAUDE.md, permissions, hooks).
- Codex: OpenAI's agent, with an open-source CLI, an IDE extension and a cloud. Included with ChatGPT, sandbox on by default.
To place these three in the wider landscape, start with our roundup of the best vibe coding tools. For the detailed reviews, read our Cursor review, our Claude Code guide and our Codex review.
The comparison table
| Criterion | Cursor | Claude Code | Codex |
|---|---|---|---|
| Company | Cursor, acquired by SpaceX | Anthropic | OpenAI |
| Main interface | Desktop editor (VS Code base) | Terminal, plus VS Code, JetBrains, desktop app, web | Terminal, plus IDE extension, desktop app, web |
| Runs inside Cursor? | — | Yes, official extension | Yes, official extension |
| Cheapest paid plan | $8/month (Plus) | $20/month (Claude Pro) | $20/month (ChatGPT Plus) |
| Models included in the plan | Cursor models (Composer, Grok) | Claude models | GPT models |
| Other models | Claude, GPT, Gemini at API price, from Premium ($30) | No | No |
| Cloud agents | Cloud Agents (from Premium) | Web sessions at claude.ai/code | Codex Cloud (from Plus) |
| Instructions file | .cursor/rules and AGENTS.md | CLAUDE.md (also reads AGENTS.md) | AGENTS.md |
| Plan mode | Plan Mode (Shift+Tab) | plan mode (Shift+Tab) | Planning in conversation |
| Built-in undo | Checkpoints | /rewind | Git commits recommended |
| Pull request review | Bugbot (from Premium) | GitHub Actions; Code Review for organizations | @codex review on GitHub (from Plus) |
| Open source | No | No | CLI open source (Apache-2.0) |
Cursor's acquisition by SpaceX was announced on August 14, 2026 on the company's blog.
What does each tier cost?
The three tools have similar price points but different contents. Monthly prices checked on November 21, 2026 on the official pages, in US dollars, before tax:
| Tier | Cursor | Claude (Claude Code) | ChatGPT (Codex) |
|---|---|---|---|
| Free | Free: limited usage, Cursor models | Free: Claude Code not included | Free: lightweight model in the desktop app, subject to rollout |
| Budget | Plus: $8 | — | Go: $8, same scope as Free |
| Standard | Premium: $30 | Pro: $20 ($17 annual) | Plus: $20 |
| Heavy | Super: $100 | Max: $100 (5x Pro) | Pro: $100 |
| Very heavy | Ultra: $200 | Ultra: $200 (10x Pro weekly) | Pro: $200 or $500 |
| Team | Teams: $30, $100 or $200 per seat | Team: $25 per standard seat ($20 annual) | Business: $25 per user ($20 annual, 2+ users) |
The trap to know: with Cursor, included usage only covers its in-house models. If you want Claude or GPT inside Cursor, you pay the subscription (Premium or above) plus usage at public API prices from prepaid credit. With Anthropic and OpenAI, the models are part of the plan, with usage limits per time window.
For a full budget (subscriptions, APIs, hosting), see what a vibe-coded app really costs.
The differences that actually matter
Interface: seeing the code or talking to it
Cursor shows you everything: file tree, files, diffs, built-in terminal. You can fix a line by hand between two requests, with Tab autocomplete. Claude Code and Codex start from the terminal: you describe, the agent acts, you review the changes. Their IDE extensions bring the two worlds closer, but the spirit stays "agent first".
If you have never opened a terminal, Cursor or the Claude desktop app will feel safer. If you like chaining commands, Claude Code and Codex are faster.
Safety defaults
- Codex starts with no network access, in a sandbox limited to the workspace. It is the most cautious out-of-the-box setting.
- Claude Code offers six permission modes, from manual to
bypassPermissions, plus anautomode where a second model checks actions. - Cursor offers three Run Modes: Auto-review, Allowlist and Run Everything.
Whatever the tool, never use the "run everything" mode outside an isolated environment. And go through the pre-launch security checklist.
Project memory
All three read persistent instructions. Good news: AGENTS.md is understood by all three (natively by Codex and Cursor, alongside CLAUDE.md for Claude Code). One well-written file lets you switch tools without re-explaining everything. The method is in our guide to CLAUDE.md, AGENTS.md and rules files.
Background work
All three can run tasks on remote machines. Cursor can also start them from Slack, Linear or an @cursor comment on GitHub. Codex Cloud keeps going while your computer sleeps. Claude Code runs in the browser at claude.ai/code and on mobile. For a solo vibe coder, that is handy for kicking off a long task (tests, a migration) and coming back later.
The decision tree
- Complete beginner wanting a visible result fast? None of the three. Start with Lovable, Bolt or Replit, then come back.
- Want a visual editor and to type some code yourself? Cursor.
- Already pay for ChatGPT Plus? Try Codex first: zero extra cost.
- Already pay for Claude Pro? Try Claude Code first, for the same reason.
- Want to switch between Claude, GPT and Gemini in one tool? Cursor Premium, watching the API bill.
- Want to automate in a script or CI pipeline? Claude Code (
claude -p) or Codex (codex exec). - Want to be able to read the tool's source? Codex CLI is open source. For other open options, see open-source vibe coding tools.
Same task, three ways to run it
To compare for yourself, give all three tools the same instruction on a small Git-tracked project. Here is the prompt to copy:
Add a /contact page with a form (name, email, message).
Validate fields on both the client and the server.
Before coding: explore the project and propose a plan of 5 steps or fewer.
After coding: run the build and tests, fix errors, then summarize the changed files.
- Cursor: open the folder, switch to Plan Mode with Shift+Tab, paste the prompt.
- Claude Code:
claude --permission-mode plan, then paste the prompt. - Codex:
codex, then paste the prompt; the sandbox is on by default.
Then compare three things: how clear the plan is, how many back-and-forths you needed and how clean the diff is. Commit before each run so you start from the same point, as explained in Git for vibe coders.
Frequently asked questions
Frequently asked questions
Can you use Claude Code or Codex inside Cursor?
Yes. Anthropic and OpenAI each publish an official extension that works in Cursor. You keep the Cursor editor, its autocomplete and interface, and run Claude Code or Codex in a panel, with usage counted against your Claude or ChatGPT plan rather than your Cursor plan.
Which one is cheapest?
For regular use with frontier models, Claude Code (Claude Pro) and Codex (ChatGPT Plus) both start at $20 a month as of November 21, 2026, models included. Cursor starts at $8, but the Plus plan is limited to its in-house models; for Claude or GPT you need Premium at $30 plus API usage.
Which is safest for a beginner?
Codex has the most cautious default: no network and writes limited to the workspace. Claude Code and Cursor also have modes that ask for your approval. In every case, real safety comes from your routine: Git, reading diffs and refusing modes that run everything unchecked.
Do you have to pick just one?
No. Many developers combine an editor and an agent, for example Cursor to write and review, and Claude Code or Codex for long tasks. A shared AGENTS.md makes the mix easier. Still, stick to two tools at first so you are not paying for three subscriptions.
Where does GitHub Copilot fit?
GitHub Copilot also has an agent mode in VS Code, JetBrains, Visual Studio, Eclipse and Xcode, plus a cloud agent that works from GitHub. It is a serious fourth option, especially if your code already lives on GitHub. We cover it in a full review of Copilot agent mode, and our general tools roundup places it against the others.
Sources
- Pricing — Cursor
- Models & Pricing — Cursor Docs
- Cursor is now a part of SpaceX — Cursor Blog
- Cloud Agents — Cursor Docs
- Run Modes — Cursor Docs
- Overview — Claude Code Docs
- Choose a permission mode — Claude Code Docs
- How Claude remembers your project — Claude Code Docs
- Plans & Pricing — Claude
- Codex IDE extension — ChatGPT docs (OpenAI)
- Agent approvals & security — ChatGPT docs (OpenAI)
- Pricing — ChatGPT docs (OpenAI)
- openai/codex — GitHub


