Cursor vs Codex: plans, prices and measured usage

Cursor is an editor with its own agent that runs models from Anthropic, OpenAI, Google and Cursor itself; Codex is OpenAI's agent, included in ChatGPT plans, and runs OpenAI's GPT models. Codex comes with ChatGPT Plus at $20 a month; Cursor's paid plans start at $8. You do not have to pick one: OpenAI's Codex extension installs in Cursor, and signed in with ChatGPT it draws on your ChatGPT plan, not on Cursor's usage.

Cursor vs Codex pricing

Cursor prices read on 20 October 2026 from cursor.com/pricing, Codex's on 10 October from OpenAI's Codex pricing.

• Cursor: now sold through The X Subscription with Grok and X. Free has limited usage; Plus is $8 a month for Grok models only; Premium is $30, Super $100 and Ultra $200. Teams seats cost $30, $100 or $200 a user; Enterprise is priced on request. Pro ($20), Pro+ ($60) and the $40 and $120 Teams seats are legacy plans, closed to new customers.
• Codex: ChatGPT Plus at $20 includes it on the web, in the CLI, the IDE extension and on iOS. ChatGPT Pro is $100, $200 or $500 a month; Business is $25 a user, or $20 billed annually, for two or more users. Free and Go include Codex in the desktop app, subject to rollout. Enterprise and Edu are priced by OpenAI's sales team.

What you get for the fee differs. Each paid Cursor plan includes a weekly pool for Cursor's own models, Grok and Composer: Premium four times Plus's, Super fifteen times, Ultra fifty. GPT models are in no plan's pool; on Premium and above they bill at OpenAI's list price from the first request. Cursor does not state the pool in dollars. A ChatGPT plan gives Codex usage limits instead, counted in messages.

Usage limits: a weekly pool against a five-hour window

Cursor's pool refreshes weekly. Past it you wait for the reset, continue on prepaid credit at each model's rates, or move to a bigger plan.

On Plus and Business, OpenAI publishes estimated local messages per five hours by model — 5 to 45 on GPT-6 Astra, 350 to 3,000 on GPT-6 Luna — and says these "are not fixed message limits". ChatGPT Pro has no five-hour limit, and weekly limits may also apply. Local and cloud work share one allowance. At the cap a running Codex turn may finish, subject to fair use, and you buy credits or switch to an API key.

What GPT models cost in each

Cursor's models and pricing page, read on 20 October 2026 with every model shown, lists GPT-5.6 Sol, Terra and Luna at OpenAI's own rates: $4, $2 and $0.20 per million input tokens, $20, $12 and $1.20 output. It lists older GPT-5 and GPT-5 Codex models too, and no GPT-6 model.

Codex runs those GPT-5.6 models and the GPT-6 line: GPT-6.1 Sol at $2 input and $10 output per million tokens, GPT-6 Astra at $10 and $50, GPT-6 Luna at $0.10 and $0.50. Above 272K input tokens OpenAI bills input at double and output at 1.5×. These are the rates Codex bills on an API key; on a ChatGPT plan, Codex's calls come out of the plan's limits.

On Cursor Teams and Enterprise, every GPT call by Cursor's agent also carries a Cursor Token Rate of $0.25 per million tokens.

Codex inside Cursor

OpenAI's Codex IDE extension installs in VS Code, Cursor and Windsurf; you open it from the Codex icon. So you can run both: Cursor's agent for quick edits on its own models, and Codex in a panel on a ChatGPT plan. Each part goes to its own bill — Cursor's agent to your Cursor plan, Codex to ChatGPT, or to OpenAI per token if you sign in with an API key.

The Codex side, measured

Caprock reads Codex's transcripts from disk and prices every call at OpenAI's list prices. It does not read Cursor's agent, so there is no Cursor column to set beside it.

Read on 3 October 2026: 16,415 Codex calls on one machine, most of them 10–28 September. A call is one request to the model — one token_count in Codex's transcript — so a turn that runs five tools is several calls. The median call cost $0.032 and re-sent 128K tokens of conversation; 97.6% of input came from the cache, and cache reads were 73% of the cost. A third of the calls ran Codex's Luna models, and 81% of tool calls were shell commands.

Cursor's agent builds its own context and picks its own models, so its calls will not match these one for one. One heavy user on one machine is not a benchmark; it is what the work cost, read the same way every time.

Which one to pick on cost

What differs is how use is counted. Cursor gives a weekly pool for its own models and bills GPT per token on top; Codex gives windows of estimated messages, and ChatGPT Pro removes the five-hour one. If you want several makers' models in one editor, or work mostly in Grok and Composer, Cursor Premium or Super fits. If you work in OpenAI's models all day, or want GPT-6, a ChatGPT plan with Codex does — and Codex runs inside Cursor if you want the editor too.

Cursor vs Claude Code and Codex vs Claude Code set each against Claude Code. The plans in full: Cursor pricing and Codex usage and cost.

See your Codex side

Inside Codex, /usage shows the account and /status the session; each sees only itself. Caprock reads every Codex session still on disk, whether or not it started it, and prices each call — per session, per repository, per model — a list-price estimate beside Codex's own windows. It does not read Cursor.

Free, open source; nothing leaves your machine. Install Caprock.