Claude Code vs Cursor vs OpenAI Codex (2026)

By July 2026, the AI coding agent market has settled into three mature, well-funded incumbents that could hardly be shaped more differently. Claude Code is a proprietary, terminal-native agent that has grown into a multi-surface product while leaving your editor alone. Cursor is the opposite bet: a proprietary IDE — historically a VS Code fork, rebuilt from scratch in version 3.0 around a multi-agent workspace — where the editor is the product. OpenAI Codex sits between them: an open-source Rust CLI plus an IDE extension, cloud tasks, and — since 9 July — a dedicated mode inside the unified ChatGPT desktop app.

All three are agentic in the full 2026 sense: they read your codebase, edit files, run commands, manage git, review PRs, and orchestrate parallel or cloud-based sub-agents. The differences that matter now are form factor, model freedom, and how the bill arrives.

The one-line answer: Claude Code if you live in the terminal and want the strongest single-vendor agent; Cursor if you want an AI-first IDE with genuine multi-model choice; Codex if you're already paying for ChatGPT or want the only open-source client of the three. The long answer, with prices and receipts, follows.

First, what happened to Codex

A widely repeated claim this month is that OpenAI killed Codex. It didn't — and getting this right matters before comparing anything. What actually happened, per OpenAI's official announcement of 9 July 2026: the standalone Codex desktop app merged into a new unified ChatGPT desktop app, where Codex is now one of three modes — Chat, Work, and Codex — available on every plan, including Free. In OpenAI's own words, "Codex remains the same powerful coding agent for developers and technical professionals," and developers can set Codex as the default view and even keep the Codex logo as the app icon.

The CLI is not merely surviving; it's on the same release train as everything else. The openai/codex repo is active (not archived, pushed 14 July), and release 0.144.4 shipped on 14 July 2026 — five days after the desktop consolidation. The product that is being retired in that announcement is Atlas, OpenAI's standalone browser. So: consolidated, not retired. If you read "Codex is dead" somewhere, it was a misreading of the desktop-app merger.

Adoption, in numbers

Read this chart honestly: GitHub stars measure repo attention, not usage. Cursor is a closed desktop product whose public repository is essentially its issue tracker, so its 33k stars badly understate a very large paid user base. The Claude Code and Codex numbers are more meaningful — both repos are where developers actually install from and file issues — but even there, only Codex's repo contains real source code. Treat the chart as a proxy for developer mindshare among terminal-agent users, nothing more.

Claude Code: the terminal agent that outgrew the terminal

Anthropic's official docs describe Claude Code as an agentic coding tool "available in your terminal, IDE, desktop app, and browser." Day to day it's cd your-project && claude, then natural-language tasks; it's Unix-composable (pipe logs in, run headless in CI), works git natively — staging, commits, PRs — and persists context via CLAUDE.md files and auto memory. The 2025 mental model of "CLI-only, macOS-only" is obsolete: there are native installers for macOS, Linux, WSL and Windows, a VS Code extension (installable into Cursor, pointedly), a JetBrains plugin, a desktop app, a web surface at claude.ai/code, with session teleport between surfaces.

Agentically it's the most orchestration-heavy of the three: subagent "agent teams" with a lead agent coordinating and merging work, background agents, Routines (scheduled runs on Anthropic-managed infra), and an Agent SDK for building your own. It also has the most candid sandbox documentation — an OS-level Bash sandbox on macOS/Linux/WSL2, with Anthropic openly noting it's unavailable on native Windows and that the default proxy does no TLS inspection.

The catches: it's Claude models only (Opus/Sonnet/Haiku/Fable — provider-flexible via Bedrock, Google Cloud, and Microsoft Foundry, but no BYOK to other models), the client is proprietary despite the public repo, and subscription usage is pooled across claude.ai chat, Desktop, and Claude Code — heavy chat use eats your coding quota.

Cursor: the IDE is the product

Cursor forked VS Code precisely so it could own the surface, and Cursor 3.0 (April 2026) rebuilt that interface from scratch "centered around agents": a sidebar showing all local and cloud agents — including ones started from mobile, web, Slack, GitHub, and Linear — an integrated browser tool, in-editor diff/PR review, and a marketplace of plugins, MCPs, and skills. It has since grown satellite surfaces (a CLI, an iOS app, Bugbot PR review), but the value concentrates in the editor.

Its standout is model choice. Cursor's models page lists Anthropic's Claude family, OpenAI's GPT-5.x line, Google Gemini, xAI's Grok 4.5 (jointly trained with Cursor), other providers, and its in-house Composer 2.5. No other tool in this trio lets you pick your model per task. BYOK exists but is second-class — on Teams/Enterprise, Cursor's token rate applies on top of API pricing even for your own keys.

The catches: proprietary end-to-end, tracks slightly older VS Code versions by its own docs, and the pricing model (two usage pools, on-demand overflow, Max Mode) is genuinely the most complicated of the three — Cursor has published clarification posts about its own pricing.

Codex: the multi-surface agent with an open-source core

Codex is OpenAI's coding agent in five current surfaces: the open-source CLI (Apache-2.0, written in Rust — the only auditable client here), an IDE extension, Codex Web cloud tasks, GitHub PR review, and the new Codex mode in the ChatGPT desktop app, with mobile access via the ChatGPT app and Codex Remote QR-pairing. Architecturally it's Claude Code's closest sibling: local and cloud tasks with handoff between hosts, sandboxed execution with approval modes, MCP support, and an AGENTS.md project-context convention.

It runs OpenAI models only — the GPT-5.6 family (Sol for maximum capability, Terra for everyday work, Luna for speed), per the official pricing docs. Auth is a ChatGPT plan or an API key. The catches: on the $20 Plus plan, cloud tasks and code reviews are simply "not available"; usage is credit-denominated with wide per-message ranges (5–40 credits), which makes budgeting fuzzier than flat quotas; and new pay-as-you-go Codex seats were discontinued for Business plans from 24 June 2026.

What they cost

Prices from claude.com/pricing, cursor.com/docs/models-and-pricing, and developers.openai.com/codex/pricing, all verified 15 July 2026.

The honest reading: these are two different billing philosophies. Claude and Codex bundle coding into a general AI subscription — your $20 buys chat and coding from one pooled quota, which is great value until a heavy week exhausts it. Cursor is closer to metered compute in a subscription wrapper: each tier includes a dollar amount of API-rate usage, then billing continues on-demand. Cursor's model is more transparent about what you consumed and more punishing if you consume a lot; OpenAI's own rate card notes Codex costs average $100–$200 per developer per month — treat every $20 entry price as an on-ramp, not a ceiling.

Which one for you

Terminal-first developers: Claude Code. It's the most mature terminal agent, the deepest orchestration stack (agent teams, Routines, headless CI mode), and the docs are refreshingly honest about limits. Codex CLI is the credible alternative — pick it if you're OpenAI-aligned or want to read the source of the tool running commands on your machine.

IDE-natives who won't leave a graphical editor: Cursor, without much contest. Neither rival is trying to be your editor — Claude Code even ships an extension that runs inside Cursor. If you want tab completions, agent panels, and PR review in one window, this is the one built for that.

Multi-model needs: Cursor is the only real option. Claude Code and Codex are single-vendor by design; Cursor lets you run Claude, GPT-5.x, Gemini, and Grok side by side and route per task.

Tight budget: ChatGPT Go at $8/mo is the cheapest paid path into a frontier coding agent, and Codex now works on ChatGPT Free for light tasks. Claude Pro at $17/mo (annual) is the strongest sub-$20 dedicated option.

Open-source requirement: none of the three qualifies fully. Codex's CLI is Apache-2.0, but the models and cloud behind it are proprietary; Claude Code and Cursor are proprietary end to end. If an auditable, self-directed stack is non-negotiable, look at OpenCode or Aider in the full roundup.

What this comparison doesn't cover

We haven't benchmarked output quality head-to-head — real-world agent performance varies by codebase, language, and prompt discipline more than any static test captures, and vendor-run evals aren't independent evidence. We also don't cover enterprise procurement details (SSO, audit, data-residency terms), which change under negotiation, or the fast-moving weekly release cadence all three ship at — specific features cited here are as of mid-July 2026. Star counts, as noted, proxy attention rather than usage, and no reliable public revenue or MAU figures exist for a cleaner comparison. Everything else was verified in July 2026 against official pricing pages, product docs, and OpenAI's 9 July consolidation announcement.