Claude Code, Cursor or Codex: which coding agent for which job in 2026?

Pick by the job, and by where you already work. For long, multi-step work in a real repo, I use Claude Code. For quick edits while I read code, an editor agent like Cursor is the fastest. For many small tasks in parallel, I send them to a cloud agent: Codex was built for that, and Cursor and Claude Code do it now too.

I use all three every day. They run side by side in tabs in my own Mac app, VShell, and my agent team is spread across them.

Claude Code vs Cursor vs Codex in one table

All three read your code, edit files, run commands and check their own work. They differ in where they live and how they like to work.

Tool	Lives in	How it works	Rules file	Models	Best for	Entry plan with the agent
Claude Code Anthropic	Terminal first; also VS Code, JetBrains, a desktop app, the web and phone	A local agent loop, with subagents and background sessions; cloud sessions too	CLAUDE.md; AGENTS.md if there is no CLAUDE.md	Claude only; Opus 5.5 by default	Long tasks across many files	Claude Pro, US$20 a month
Cursor now owned by SpaceX	An editor (built on VS Code) first; also a CLI, web and phone	An editor with agents, many in parallel on your machine or in cloud VMs	.cursor/rules/*.mdc and AGENTS.md	Its own Composer 2.5 and Grok 4.7; Claude, GPT and Gemini on higher plans	Edits while you read and steer	Cursor Plus, US$8 a month, Grok models only; Premium US$30
Codex OpenAI	Cloud tasks, a CLI and an editor extension; the desktop app is now inside ChatGPT	A local agent in a sandbox, or one cloud workspace per task	AGENTS.md, up to 32 KiB	OpenAI only; GPT-6.1 Sol for hard work	Many small tasks in parallel	ChatGPT Plus, US$20 a month

For me the rules file column matters most. All three read AGENTS.md now, so one file can serve a team that uses two tools. How to write it is in my article on a good CLAUDE.md.

Which coding agent for which job

Job	Claude Code	Cursor	Codex
Big refactor, long task	subagents, 1M context	Projects (beta)	worktrees, subagents
Quick edits while you read	VS Code extension	built for it	IDE extension
Many small tasks in the cloud	cloud sessions, GitHub only	cloud VMs, API prices	one workspace per task
Review a pull request	Team and Enterprise	Bugbot	@codex review
Learn a new codebase	pick where you read	pick where you read	pick where you read

My marks from daily use, checked against each tool’s docs, October 2026.

These are my marks from daily use, checked against each tool’s docs. Three dots: the one I reach for first. Two dots: it does the job well, but it is not my first pick.

Big refactor or a long, multi-step task

I reach for Claude Code. It works in my real repo and can split the work. A lead agent hands parts to subagents, each with its own context and, if I ask, its own copy of the repo in a git worktree. Opus 5.5 takes up to 1 million tokens of context, and Anthropic says its Fable models are “suited to tasks larger than a single sitting”.

Watch out: subagents draw on the same usage limit as the main session. A big refactor can use up a five-hour window.

Quick edits while you read code

I reach for the editor. To read a file, change three lines and see the diff right there, an editor agent is faster. Cursor is built for this. Codex and Claude Code have VS Code extensions too, and OpenAI’s own advice is to “keep quick iterations local”.

Watch out: on Cursor’s new US$8 Plus plan you get Grok models only. Claude and GPT models start at Premium, and you pay their API list price from the first request.

Many small tasks in parallel, in the cloud

Any of the three can do it now. Codex started in May 2025 as a cloud agent with one sandbox per task, and each task “can keep working while your computer is asleep”. Cursor runs “as many agents as you want in parallel” in cloud VMs. Claude Code’s cloud sessions keep running after you close your laptop.

The differences are cost and hosting. Cursor charges cloud agents at API prices and asks you to set a spend limit first. Claude’s cloud sessions need GitHub, while Cursor also clones from GitLab, Azure DevOps and Bitbucket. I give the cloud only small, clear tasks with a test that proves them, like “add the missing test for the refund rule”.

Review of a pull request

Codex or Cursor’s Bugbot, then a person. Codex review on GitHub is generally available: write “@codex review” in the pull request. Bugbot runs on every update of a pull request, on GitHub, GitLab, Bitbucket and Azure DevOps. Claude’s managed Code Review is a research preview for Team and Enterprise, at US$15 to 25 per review on average, billed on top of the plan. On any plan you can mention @claude in a pull request through its GitHub Action.

None of these bots approve or block your merge by default. That part stays with a person. How I check AI-written code is in my article on checking AI code quality.

Learning a new codebase

All three are good here. Pick the one where you like to read. None of them builds a vector index of your code now. They search it with fast text search, much like you would with grep. Cursor stopped storing embeddings for search in July 2026 (its docs). I like to ask for a map of the modules and the path of one request, with the agent reading only. More in my article on RAG in daily coding.

Honest limits of each, from their own docs

Claude Code

• from the docsClaude models only
• from the docsOne usage pool with Claude chat: 5-hour window + weekly limit
• from the docsAuto mode is the default since 14 Aug 2026
• from the docsCloud sessions need GitHub

Cursor

• from the docsPlus plan: Grok models only
• from the docsClaude and GPT at API prices, Premium and up
• from the docsCloud agents billed at API prices
• from the docsTurn Privacy Mode on

Codex

• from the docsOpenAI models only
• from the docsModels change fast (GPT-5.5 left on 14 Oct 2026)
• from the docsPlus: 15 to 160 Sol messages per 5 hours
• from the docsNetwork off by default

Claude Code. It runs Claude models only. Claude chat and Claude Code share one usage pool, with a five-hour window and a weekly limit on paid plans. Since 14 August 2026, auto mode is the default for new sessions on Pro, Max and Team. In auto mode, “a classifier reviews actions instead of you”. It is faster. Know that it is on.

Cursor. SpaceX bought Cursor, and the deal closed on 14 August 2026. Cursor is now sold inside “The X Subscription” with Grok and X. The old Pro and Teams plans are “legacy” and keep working. With Privacy Mode off, Cursor “may use and store codebase data” to train its models. Turn it on.

Codex. It runs OpenAI models only, and they change fast: GPT-5.5 left Codex on 14 October 2026. On the Plus plan you get 15 to 160 local messages with GPT-6.1 Sol per five hours, and “cloud tasks may use more of your allowance”. The network is off by default. That is safe, but a task that must install packages needs it on.

In daily use, the limit I hit most is usage. When one account runs low in the middle of a task, I move the task to another account or another tool.

What matters more than the tool

Rules file

Tests it can run

the agent

Small tasks

Your review

A 2026 study of 7,156 agent pull requests found that the kind of task mattered a lot: 82.1% of documentation changes were accepted, against 66.1% of new features. In the authors’ words, “No single agent performs best across all task types”.

Two Carnegie Mellon studies of open-source projects found that the speed boost after a team adopts Cursor or an agent fades, while linter warnings and code complexity stay higher. Tests and review keep that in check. Boris Cherny, who created Claude Code, said that giving the agent a way to verify its work will “2-3x the quality of the final result” (quoted by paddo.dev).

So four things decide most of my results, whatever the tool: a short rules file, tests the agent can run, tasks small enough to review in one sitting, and a person who reads the diff. Setting this up for a team’s agents is part of what my team does; here is how we work.

Many teams use two of them

1. Editor agentread, steer, small edits
2. bigger than a quick editTerminal agentthe long task, on a branch
3. split into small ticketsCloud agentssmall tasks in parallel

Cloud back to editor: review the diff

Shared: AGENTS.md, tests, CI

In The Pragmatic Engineer’s survey of 906 engineers (February 2026), 70% used two to four AI tools and only 15% used one. Andrej Karpathy described his setup in January 2026: “a small few CC sessions on the left” (CC is Claude Code) “and an IDE on the right for viewing the code + manual edits”.

In my agent team, the developer agent runs on Claude Code and the research agent on Codex. One rule makes the mix work: each agent gets its own branch and its own part of the code. In a July 2026 study of pull requests open at the same time, two from the same agent clashed in 19.8% of pairs, and two from different agents in 41.7%.

How many developers use each one: 2025, 2026 and a guess for 2027

Coding agents used at workShare of professional developers, JetBrains surveys. The 2027 zone is a guess, with no numbers.

• Claude Code
• Codex
• Cursor
• GitHub Copilot

Source: JetBrains Research, AI Pulse surveys (Apr 2026, over 10,000 developers; Aug 2026, over 15,000 developers). Use at work. JetBrains gives no 2025 figure for Codex, Cursor or Copilot.

• 2025. About 3% of professional developers used Claude Code at work in April to June 2025.
• 2026. JetBrains asked more than 15,000 developers in May to July 2026. At work, 39% used Claude Code, 21% GitHub Copilot, 16% Codex and 12% Cursor, and 90% used a coding agent at least once a week. Stack Overflow asked people who use AI coding tools which agents they used in the past year: Claude Code 65.5%, Copilot 58.7%, Codex 29.5%, Cursor 25.5% (published 6 October 2026).
• 2027 (my guess). Two tools per developer becomes normal. Claude Code stays first and Codex keeps growing fastest: 17.6% of Stack Overflow’s respondents want to use Codex in the next six months, close to Copilot’s 19.6%. Cursor’s place depends on how the new X Subscription plans suit developers.

The two surveys ask different people different questions, so the levels differ. The order is the same in both: Claude Code, Copilot, Codex, Cursor.

How to choose in 10 minutes

1. Minutes 0 to 2: where do you work all day? In the terminal, start with Claude Code or the Codex CLI. In an editor, start with Cursor, or the Codex or Claude Code extension in VS Code.
2. Minutes 2 to 4: which models do you want? Claude Code runs Claude. Codex runs OpenAI. Cursor runs its own models and Grok, and others at API prices on Premium and up.
3. Minutes 4 to 7: check the rules for your code. Where is it hosted? Claude’s cloud sessions need GitHub. Must your code stay out of training? Use a business plan, or Privacy Mode in Cursor.
4. Minutes 7 to 10: give one real ticket to two of them. Same AGENTS.md, same task, separate branches. Compare the diff and the test run, not the chat.

vaisoftlab.com