Claude Code vs Codex vs Cursor (an honest comparison)

TL;DR

• Claude Code is a terminal-native adoption machine — Theo says Anthropic won by meeting developers in the terminal, helping Claude Code replace Cursor in YC-style startup workflows, but that convenience comes with UI compromises and a product tuned for demos and screenshots.

• Anthropic optimizes for feeling productive, often by burning more tokens — features like sub-agents, pet mode, and flashy terminal UI create a strong sense of momentum, but Theo’s core critique is that Claude Code repeatedly solves problems by spending more tokens rather than finding more efficient verification loops.

• Codex is intentionally boring in the best way — compared side by side, Codex shows far less UI activity, but Theo argues OpenAI is focused on practical features like locked-Mac computer use, screenshot context capture, and model efficiency, especially with GPT-5.5 using roughly half the tokens of peers on benchmarks while scoring better.

• OpenAI dogfoods the same Codex app users get; Anthropic largely does not — Theo claims thousands of OpenAI employees, including non-engineers, use the public Codex app every day, while Anthropic employees use internal Claude tooling, different prompts, and the unreleased Mythos model, which helps explain rough edges in the public Claude desktop app.

• Cursor’s cloud agents are the sleeper pick — Theo says Cursor fell from first to third mostly because people missed its strongest product: cloud sandboxes with a full graphical Linux environment that can run apps, use computer vision, post Slack-thread fixes, and even return proof via video.

• His practical buyer’s guide is brutally simple — Claude Code fits devs who want motivation and visible momentum, Codex fits experienced engineers who want reliable workflows that stay out of the way, and Cursor fits teams that want enterprise-ready cloud agents non-technical coworkers can trigger from tools like Slack.

The Breakdown