Cursor vs Codex: an AI editor or a console agent, which to choose

Cursor vs Codex is a comparison not of two models but of two ways of working. Cursor is a code editor with built-in AI, where you see files, diffs and chat side by side. Codex CLI is OpenAI's open-source console agent: you give it a task in the terminal, and it reads files, runs commands and makes edits. Below we compare workflows, bring-your-own-provider setup and API cost so you can decide between cursor or codex for your work. We do not quote subscription prices for either service: they change, so look at the current ones on the Cursor and OpenAI websites.

Cursor or Codex: the fundamental difference

Two caveats. First, both products keep evolving: Cursor has an agent mode too, and Codex can be built into an IDE workflow, so the table describes typical use, not hard boundaries. Second, we do not judge whose agent writes better code: quality depends on the chosen model and task, so check on your own examples.

Workflows: when each is more convenient

Cursor fits if you read and edit a lot of code by hand and want AI at your fingertips: select a function, ask for a rewrite, look at the diff, accept. It is convenient when you need to see context, attach files and folders to the chat and switch quickly between models.

Codex fits if a task can be stated whole and checked with commands: "add an endpoint, update the tests, run them". A terminal agent is easy to run on a remote server over SSH, in CI, in several windows in parallel. Project rules go in AGENTS.md.

Together: nothing stops you from keeping an editor for reading and manual edits and handing long autonomous tasks to a terminal agent. The bill stays shared: both tools can work with one key and one balance.

Connecting your own provider

Both tools can be pointed at the Zukko gateway. One fetch_... key and one dollar balance work in both.

Cursor

Per the Zukko documentation for Cursor↗: in Cursor Settings → Models turn on OpenAI API Key and paste the key, turn on Override OpenAI Base URL and enter https://api.zukko.pro/v1, add a model and press Verify. Per the Zukko documentation, overriding the provider is not available on Cursor's free plan; a paid plan is needed.

An important detail: Cursor does not let you add a custom model whose name it already knows, and it will send such a request to its own backend. That is why popular models have second names, for example cursor-sonnet-5 instead of claude-sonnet-5 and cursor-opus-55 instead of claude-opus-5-5. The full table of names is in the documentation. If something does not work, see the article Cursor does not see models↗.

Codex

Codex reads ~/.codex/config.toml and ~/.codex/auth.json (per the Zukko documentation for Codex↗):

model = "gpt-6.1-sol"
model_provider = "zukko"
model_reasoning_effort = "high"

[model_providers.zukko]
name = "zukko"
base_url = "https://api.zukko.pro/v1"
wire_api = "responses"
requires_openai_auth = true

{ "OPENAI_API_KEY": "fetch_your_key" }

You can change the model in the config or at launch: codex -m <model name>. Codex talks to Zukko through the Responses API, so GPT and Claude models fit; GLM models do not work in it. There are ready installers for macOS and Windows, described on the documentation page.

About official subscriptions (ChatGPT for Codex, Cursor plans) we do not know and do not quote prices or terms: see the current plan on the service's website.

What it costs through the API

When you work with a key you pay for tokens. Formula without cache: cost = input × input price / 1,000,000 + output × output price / 1,000,000. Zukko prices as of October 18, 2026, $ per 1M tokens:

The index is the Intelligence Index from Artificial Analysis; we give no other quality ratings.

Assumption for the example: 100 agent requests, each with 40,000 input tokens and 2,000 output tokens. Input 100 × 40,000 = 4M, output 100 × 2,000 = 0.2M.

• gpt-6.1-sol ignoring cache: 4 × $0.27 + 0.2 × $1.32 = $1.08 + $0.264 = $1.344.
• The same if 75% of input is read from cache (an assumption): 1M fresh, 3M from cache. 1 × $0.27 + 3 × $0.013 + 0.2 × $1.32 = $0.27 + $0.039 + $0.264 = $0.573. For GPT models the gateway sees cache reads in both formats.
• claude-sonnet-5 in Cursor (there it is added as cursor-sonnet-5; OpenAI format, Claude cache not visible, all input at the input price): 4 × $0.4 + 0.2 × $2 = $1.60 + $0.40 = $2.00. The same tokens at official prices: 4 × $2 + 0.2 × $10 = $8 + $2 = $10.

A practical conclusion about format follows. For Claude, cache gives a discount only in the Anthropic format (/v1/messages, Claude Code, Anthropic SDK). Cursor and Codex work in OpenAI formats, so in them it makes sense to look at GPT models, whose cache is visible, and to count Claude at the full input price. Details are on the key limits and cache↗ page. See your real numbers in the "Usage" section of the dashboard, and current prices in GET https://api.zukko.pro/v1/models.

How to choose in one evening

1. Pick one real half-day task: a small feature with tests.
2. Do it in Cursor and in Codex on the same model.
3. Compare the time, the number of manual fixes and the actual spend in the dashboard.
4. Repeat on a second model from the table.

For safety, create a separate key with a daily dollar limit on the API keys↗ page: the agent cannot run you into the red if it loops. The balance is topped up from $1: SBP (the payment page adds 8.5% on top), crypto or a MasterCard card.

FAQ

Can Cursor and Codex use the same key?

Yes. One fetch_... key and one balance work for both. The address is https://api.zukko.pro/v1 in both cases, but Cursor needs second names for some models.

Which is cheaper?

The tool itself does not set the token price: it depends on the model and on whether cache is visible. Calculate with the formula above on your own profile.

Do I need a Cursor subscription for my own key?

Per Zukko documentation, overriding the provider requires a paid Cursor plan. Check plan terms on the Cursor website.

Which tool is better for a beginner?

If you are used to an editor and want to see diffs, start with Cursor. If you are comfortable in the terminal, try Codex. Your workflow decides, not abstract quality.