Cursor vs Codex: which one to use after Lovable

Two things changed in October 2026. On 12 October new Cursor customers started on The X Subscription, where GPT, Claude and Gemini are not part of the plan. And OpenAI says it plans to stop supplying its models to Cursor, with 12 November 2026 as the proposed date. Most "Cursor vs Codex" articles were written before that.

We read the official docs, help pages and pricing pages of both on 26 October 2026. We did not run our own benchmark.

Short answer

• Pick Cursor if you want an editor where you see every file and undo the agent with one click, and Cursor's own models (Grok and Composer) or Claude at API prices are fine for you.
• Pick Codex if you want OpenAI models for a fixed monthly price, you already pay for ChatGPT or you would rather give tasks to an agent than work in an editor.
• Want both? Use Cursor as the editor and run Codex inside it on your ChatGPT plan.

Side by side

Prices checked on 26 October 2026, in US dollars a month, without tax. For Cursor these are the plans new customers get. Both change often, so check the pricing pages before you pay.

What you pay

Codex comes with every ChatGPT plan, also the free one. But on Free and Go ($8) the pricing page lists only GPT-6 Luna in the desktop app. Codex in the terminal, the editor extension and on the web starts at Plus, $20 a month, with GPT-6.1 Sol. That is the model OpenAI recommends for complex coding. Pro costs $100, $200 or $500 a month.

Cursor has more steps:

• Free, $0: limited Agent and Chat. No third-party models.
• Plus, $8 a month or $80 a year: Cursor's own models only, at fixed settings. No Claude, no GPT, no own API keys and no Cloud Agents.
• Premium, $30 a month or $300 a year: 4x the Plus usage. Adds Cloud Agents, Bugbot and third-party models (Claude, GPT, Gemini) at API list price, paid from prepaid credit.
• Super, $100, and Ultra, $200: 15x and 50x the Plus usage. Third-party models are still billed by usage on top.

So if GPT is the model you want, Codex is the cheaper way to get it. On Cursor you pay $30 for Premium and then pay for every GPT request on top. If Cursor's own models do the job, Cursor Plus at $8 is the cheapest paid plan of the two with an agent in an editor.

Already paying for Cursor? Your plan keeps working. Pro, Pro+, Ultra and the old Teams seats are now legacy plans with their own monthly pools. Switching to The X Subscription is optional, but Cursor's help page says it "can't be undone" and Support cannot undo it for you. If you use GPT or Claude in Cursor, look at your usage before you switch.

The OpenAI cutoff

OpenAI's help center says it is "planning to wind down our contract providing OpenAI models to Cursor". It has proposed 12 November 2026 as the date. Until then Cursor keeps access to the OpenAI models it uses today. OpenAI also writes that "Cursor may choose to end access sooner" and that it will share the official date once both companies confirm it. The page was updated on 25 October 2026.

On 26 October Cursor's model list had no GPT-6 models. The newest OpenAI models there were GPT-5.6 Luna, Sol and Terra. A ChatGPT plan does not pay for GPT in Cursor's own Agent.

To keep using GPT inside Cursor, OpenAI names three options:

• Your own OpenAI API key, in Cursor Settings > Models. Billed at OpenAI API prices, separate from ChatGPT. It works only in local Chat and Agent, not in Tab, Auto, Cloud Agents, Automations or the Cursor CLI. Cursor's docs list OpenAI key support as "Standard, non-reasoning chat models", and Plus does not allow own keys.
• The Codex extension inside Cursor, signed in with ChatGPT or an API key. It does not use Cursor's model list.
• A gateway like Amazon Bedrock or Azure, with the same limits as your own key.

If you use Cursor for Grok, Composer or Claude, not much changes for you. If GPT is the reason you use Cursor, the Codex extension is the simple option. Whether the cutoff happens on 12 November or another day, these three options stay the same.

What happens when you hit a limit

Cursor gives you one usage pool that refreshes every week. The editor, the CLI, Cloud Agents and the Grok apps all draw from it. When it runs out, you wait for the reset, pay with prepaid credit or move to a higher plan. You see what is left in the editor settings or at grok.com/dashboard. GPT and Claude do not use the pool. They come out of prepaid credit, so set a monthly limit for extra usage before your first request.

Codex on Plus counts usage per five hours, and weekly limits may apply too. OpenAI gives estimates, not fixed numbers: 15 to 160 local messages per five hours with GPT-6.1 Sol and 350 to 3,000 with GPT-6 Luna. Cloud tasks use more. Pro plans have no five-hour limit right now. When you hit the limit in the middle of a task, Codex can finish the turn it is on. Then you wait, buy ChatGPT credits or switch to an API key. In the terminal, /status shows what is left.

Codex also shares its usage with ChatGPT Work. If you use that a lot, it comes out of the same budget.

How much it does without asking

Here the two start from different places, and for a Lovable app it matters. Your Supabase project is live, with real users in it.

Codex locks itself in. On your computer it runs in a sandbox enforced by the operating system, with no network and writes limited to the project folder. In the Auto preset it reads, edits and runs commands in that folder on its own. It asks before it edits anything outside the folder or runs a command that needs the network. A command that reaches your Supabase project needs the network, so Codex stops and asks you. /permissions switches it to read-only when you only want to plan.

Cursor has three Run Modes, in Settings > Agents > Approvals & Execution:

• Auto-review: commands on your allowlist run straight away, other shell commands run in a sandbox when they can and the rest go to a classifier that decides.
• Allowlist: only the actions on your list run without asking.
• Run Everything: no prompts, no sandbox and no classifier.

Cursor calls Auto-review "the safest useful setup for most people" and in the same docs says it "is not a security boundary". The sandbox blocks the network apart from package managers and the domains you allow. So a command that changes your live database cannot run in the sandbox. It goes to the classifier, not to you, and the classifier can make mistakes.

So for the first days: in Cursor, use Allowlist, or add a block instruction for database commands in .cursor/permissions.json. In Codex, keep the default and read what it asks. In both, back up the database before the agent touches anything. Back up the database of a Lovable app shows how.

Editor or task list

Cursor is an editor. You have the file tree, every file opens in a tab and you see the edits land. Before big changes the agent makes checkpoints, and you click one to put the files back. Checkpoints are separate from Git, so they also work if you have never made a commit. Tab also completes code as you type, if you want to change code yourself.

Codex is an agent you give tasks to. You describe the change, it works and you read the result as a diff. In the editor extension you keep or undo each edit. Undo beyond that goes through Git, and OpenAI suggests a commit before and after each task. /review in the terminal checks your changes and reports problems.

If you have never opened a code editor, Cursor's checkpoints are the easier safety net. Lovable vs Cursor and Lovable vs Codex compare each one with Lovable.

Notes for the agent

Both read AGENTS.md, a plain Markdown file in the project. Codex reads it before it does any work, up to 32 KiB of these files combined by default. /init in the Codex terminal writes a first version for you.

Cursor reads AGENTS.md too. For more control it has project rules in .cursor/rules. They must be .mdc files, and a plain .md file there is ignored. /create-rule in Agent writes one for you. Cursor recommends keeping rules under 500 lines.

So one AGENTS.md works for both, and your notes do not lock you into either tool. Take your Lovable Knowledge to CLAUDE.md or AGENTS.md shows how to bring along what you wrote in Lovable.

Which writes better code?

For a new Cursor customer on Plus, "Cursor vs Codex" is mostly "Grok and Composer vs GPT-6.1 Sol". Most benchmarks you will find tested older models. OpenAI alone retired three models from Codex between August and October 2026: GPT-5.4, GPT-5.3-Codex-Spark and GPT-5.5.

We are not going to pick a winner from benchmarks that are out of date in a month. For a typical Lovable app (React, Supabase, a few edge functions), give one real task to the model you would actually pay for, and compare the result and the bill. Your code is a normal Git repository, so switching later costs you an install and a sign-in.

Can you use both?

Yes, and it is one of the options OpenAI itself points to. Install the Codex extension in Cursor from the Codex IDE extension page, open Codex from the Activity Bar and choose Sign in with ChatGPT. It runs its own agent on your ChatGPT plan and does not use Cursor's model list. You get Cursor's editor and checkpoints, and GPT-6.1 Sol on ChatGPT Plus. The Codex pricing page lists the editor extension from Plus up.

If you do this, do not let Cursor's agent and Codex edit at the same time, and commit often. If Lovable is still connected to the same repository, read Keep Lovable in sync when you edit in another tool first.

Getting started after Lovable

• Connect your Lovable project to GitHub and get the repository onto your computer. Export a Lovable project to GitHub and Run a Lovable project on your computer show how.
• Cursor: download the editor from cursor.com/download, open the project folder and start Agent with Cmd+I. Prefer the terminal? Install the Cursor CLI and run agent in the project folder.
• Codex: install the CLI, run codex in the project folder and choose Sign in with ChatGPT. Or use the Codex extension in your editor.
• Give the agent its notes: run /init in Codex, or ask Cursor's agent to write an AGENTS.md.
• Pick a safe mode before the first real task: Allowlist in Cursor, the default sandbox in Codex. Commit before you start.

Cursor CLI on macOS, Linux or WSL:

curl https://cursor.com/install -fsS | bash

Codex CLI on macOS or Linux:

curl -fsSL https://chatgpt.com/codex/install.sh | sh

The Lovable to Cursor and Lovable to Codex guides go further: database, logins, edge functions, keys, hosting and the domain. They ask a few questions about your app and show only the steps you need.

Pick Cursor if, pick Codex if

Pick Cursor if you want to see every file in an editor and undo the agent with a checkpoint, you also write code yourself or Cursor's own models or Claude do the job. Start on Plus at $8 if Grok and Composer are enough. Go for Premium only if you also want Claude or GPT by usage, and remember OpenAI plans to end GPT in Cursor's own model list.

Pick Codex if you want OpenAI models for a fixed price, you already pay for ChatGPT or you are happy to give tasks to an agent and check the diff. Start on Plus at $20. Go for Pro if the five-hour limit gets in your way. And if you still want an editor, run Codex inside Cursor.

Sources

1. Cursor: Pricing
2. Cursor docs: Models & Pricing
3. Cursor help: What is The X Subscription?
4. xAI: The X Subscription
5. Cursor docs: API keys
6. Cursor docs: Run Modes
7. Cursor docs: Agent
8. Cursor docs: Rules
9. Cursor docs: CLI
10. Cursor: Download
11. OpenAI Help Center: Using OpenAI models in Cursor
12. Codex docs: Pricing
13. Codex docs: Models
14. Codex docs: CLI
15. Codex docs: IDE extension
16. Codex docs: Approvals and security
17. Codex docs: AGENTS.md

• Lovable to Codex
• Lovable to Cursor
• Lovable vs Cursor
• Lovable vs Codex
• Back up the database of a Lovable app
• Take your Lovable Knowledge to CLAUDE.md or AGENTS.md
• Codex vs Claude Code