Compare

Codex vs Cursor.

OpenAI’s agent on a ChatGPT plan against Cursor on the X Subscription. Cursor’s new plans include its own Grok and Composer models, GPT models there cost their API price, and Cursor’s list stops at GPT-5.6. Every figure below was read on 16 November 2026 from the page linked under Sources. Starkomand, which runs both CLIs, is ours; the comparison does not depend on it.

v1.0.8 · Mac and Windows · free for 3 days · no account

The short version.

  • GPT modelsCodex runs GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Sol and GPT-6 Luna on a ChatGPT plan. Cursor lists GPT-5.6 Sol, Terra and Luna at API list price from prepaid credit, and no GPT-6 model.
  • PriceBoth have an $8 tier: ChatGPT Go for light Codex tasks and Cursor Plus for Grok and Composer. Codex’s full plan is ChatGPT Plus at $20; Cursor with third-party models starts at Premium, $30.
  • BenchmarksCodex CLI with GPT-6 Astra has the top Terminal-Bench 2.1 run, 87.4%. Cursor CLI with Grok 4.5 has 79.3%.
  • DefaultsCodex starts inside an OS-enforced sandbox with network access off. Cursor CLI asks before commands and writes that are not on its allow list.
  • InstructionsBoth read AGENTS.md, so one file of rules covers them.

Side by side.

Codex CLI next to Cursor, from both vendors’ public pages on 16 November 2026.

Criterion Codex CLIOpenAI CursorAnysphere, part of SpaceXAI
Made by OpenAI Anysphere, a SpaceX subsidiary since August 2026
Where it runs The terminal, IDEs, the desktop app, the web and iOS, plus cloud tasks The Cursor editor, its CLI, cloud agents, the web and mobile
Models in the plan GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Sol and GPT-6 LunaEach model draws on the allowance at its own rate. Grok 4.7, Grok 4.6, Grok 4.5 and Composer 2.5One pool that refreshes weekly, shared with the Grok apps.
GPT models Included in the plan GPT-5.6 Sol, Terra and Luna at API list price, on Premium and upNo GPT-6 model is on Cursor’s list.
Terminal-Bench 2.1, best published run 87.4% with GPT-6 Astra 79.3% with Grok 4.5
OS sandbox On by default, with network access offWrites limited to the working directory. Not on by default; allow and deny lists set what runs without asking
Project instructions AGENTS.md .cursor/rules and AGENTS.md
Cheapest plan with the agent ChatGPT Free and Go ($8/month) for light tasks; Plus at $20/month Free, with limited agent requests; Plus at $8/monthPlus covers Cursor’s own models only.
Top individual plan ChatGPT Pro, $500/monthPro also comes at $100 and $200. Ultra, $200/month50 times Plus usage.
Usage windows Five-hour windows on Plus; none on Pro at present; weekly limits may apply One pool that refreshes weekly

Same on both. MCP servers, skills, subagents and cloud agents that run on the vendor’s machines.

Every figure on this page was read on 16 November 2026 from the page listed under Sources, and both products change monthly; product names belong to their owners. Spotted a change? Tell us.

Benchmarks.

Terminal-Bench measures agents completing real command-line tasks. Version 2.1 fixes 28 of the 89 tasks in 2.0, and no task in 2.1 is left unsolved.

Harness Model Score Run date
Codex CLI GPT-6 Astra 87.4% 3 Sep 2026
Codex CLI GPT-5.5 83.2% 23 Apr 2026
Cursor CLI Grok 4.5 79.3% 8 Jul 2026
Codex CLI GPT-5.6 Terra 78.4% 26 Jun 2026
Codex CLI GPT-5.6 Luna 75.7% 26 Jun 2026

Selected rows from the Terminal-Bench 2.1 leaderboard.

Cursor’s run uses its own model
Cursor CLI’s 79.3% is with Grok 4.5, a model its plans include. No run puts a GPT model in Cursor CLI.
The newest models are missing
GPT-6.1 Sol, GPT-6.1 Terra, Grok 4.7 and Composer 2.5 have no runs on the board yet.

No benchmark measures your repository. The better test is one real task run in both: note the wall-clock time, what it drew from your allowance and whether the diff shipped. It is a sample of one, so repeat it after the next model release.

Prices.

OpenAI sells Codex inside ChatGPT plans; Cursor sells a weekly pool inside the X Subscription.

Plan Price What it includes
ChatGPT Free $0 Quick coding tasks with GPT-6 Luna in the desktop app, subject to rollout
ChatGPT Go $8/month Lightweight coding tasks with GPT-6 Luna in the desktop app, subject to rollout
ChatGPT Plus $20/month Codex on the web, CLI, IDE and iOS; GPT-6.1 Sol and GPT-6 Luna, with per-model allowances
ChatGPT Pro $100, $200 or $500/month No five-hour limit at present; Ultrafast mode on the $500 plan
Cursor Free $0 Limited Agent and Chat usage; no third-party models
Cursor Plus $8/month Grok 4.7, 4.6, 4.5 and Composer 2.5 at fixed settings; no Claude, GPT or Gemini
Cursor Premium $30/month Four times Plus usage, cloud agents, and Claude, GPT and Gemini at API list price from prepaid credit
Cursor Super $100/month 15 times Plus usage
Cursor Ultra $200/month 50 times Plus usage

On Cursor Premium a GPT-5.6 Sol request costs $4 in and $20 out per million tokens from prepaid credit. On a ChatGPT plan, Codex’s newer GPT-6.1 Sol draws on the plan’s allowance instead.

Which one.

Pick Codex if

you want GPT-6 models on a flat plan, an OS-enforced sandbox from the first run, or the top published Terminal-Bench run.

Pick Cursor if

you want an editor with the agent built in, Grok and Composer without a per-token bill, or one subscription that also covers Grok and X.

Pick both if

you edit in Cursor and want Codex for long agent runs. ChatGPT Plus and Cursor Plus together cost $28 a month, and one AGENTS.md serves both.

MoreCodex CLI in StarkomandCursor in StarkomandCodex vs Claude CodeCursor vs Claude CodeOpenCode vs Claude CodeGitHub Copilot vs Claude CodeGemini CLI vs Claude CodeRun multiple Claude Code agents in parallelWhat is an ADE?Agent orchestrationcmux for Windows

Running both by hand.

Both read AGENTS.md, so one file of rules covers them. Install Cursor’s CLI and run each agent in its own terminal.

# Cursor CLI: installs the agent command (cursor-agent works too)
curl https://cursor.com/install -fsS | bash

# a worktree each when both write code
git worktree add ../app-cursor -b cursor-task

# then, in a terminal each
codex
cd ../app-cursor && agent

Give the two agents different jobs, and when both write code, give each its own git worktree so neither overwrites the other’s changes.

Running both in Starkomand.

Starkomand is ours, so this is the one section about it. It runs on Apple silicon Macs and on Windows 10 and 11, and needs no Starkomand account: Codex and Cursor’s CLI each run in a real terminal window on one canvas, signed in with your own plans or keys.

Plan in one, build in the other
Open Agent actions in a window, choose Hand off… and pick the other agent. The brief carries the objective and acceptance criteria, and the new window lands wired to the one that wrote it.
See both allowances
Settings → Usage meters puts a ring in the rail for your ChatGPT plan and one for Cursor, each with its reset.
Read across windows
An agent can read another window’s transcript by name, so the plan does not need retyping.
Redirect by voice
Hold Option and speak to an agent by name.

We have not measured whether this makes anyone faster. What it changes is that each agent stays on its own plan, in view, and work moves between them without retyping.

Questions

Is Codex better than Cursor?

On the published benchmark, Codex CLI with GPT-6 Astra has 87.4% on Terminal-Bench 2.1 against 79.3% for Cursor CLI with Grok 4.5. Those are different models, so run one real task in each before you decide.

Can I use GPT models in Cursor?

On Premium ($30 a month) and higher, at API list price from prepaid credit. Cursor’s list on 16 November 2026 offered GPT-5.6 Sol, Terra and Luna and no GPT-6 model.

Which is cheaper, Codex or Cursor?

Both start at $8: ChatGPT Go for light Codex tasks and Cursor Plus for Grok and Composer. For GPT-6 models, Codex on ChatGPT Plus at $20 is the way in; Cursor charges per token for every model outside its own.

Do Codex and Cursor read the same instructions file?

Yes. Both read AGENTS.md in the project root, and Cursor also reads its own .cursor/rules.

Can I use both in one project?

Yes. Keep the rules in AGENTS.md and give each agent its own git worktree when both write code.

Try it against your own setup.

Every feature, every provider, free for 3 days. Your CLIs are already installed; the canvas takes a minute.

v1.0.8 · Mac and Windows · no account