Claude Code vs Codex vs Cursor: a philosophy comparison, not a benchmark

2026-05-26 Tue ⏱ 37 min @t3dotgg

https://www.youtube.com/watch?v=JMYspR42HFM
Original title: Claude Code vs Codex vs Cursor (an honest comparison)

Theo compares the three main coding agents not by model smarts but by philosophy — what each team is optimizing for and where they're betting the future is. Claims they differ far more than people think.

Claude Code — the "slot machine"

Built to meet devs in their terminal; easy install, no workflow change. Took the mental space Cursor used to own.
Opus 4.5 (late last year) was the "oh sh*t" moment that made trusting a terminal agent worth it; Theo now runs everything in bypass/yolo mode.
As much a marketing tool as a dev tool — features optimized for the Twitter screenshot (pet mode, sub-agents, lo-fi radio). Engineered to feel productive ("gamified"). Philosophy: when more tokens can solve it, use more tokens.
Anthropic discourages programmatic CLI use / external integrations (now charges much more) — wants lock-in inside their CLI. Desktop app is badly neglected (Theo hits login failures live).

Codex — boring and productive

Minimal UI, no flashy animations — optimized to get work done, not to demo. Ships practical paper-cuts (computer use even when Mac is locked, diff markers, screenshot hotkey).
OpenAI is pushing token efficiency (GPT-5.5 better score at half the tokens) and verification via computer use rather than spinning up 15 checker agents.
Betting on today: make agents succeed now. Uses your existing local machine config rather than fighting cloud environments.

Cursor — betting furthest out

Real power is now the cloud, not the IDE. Cloud agents spin up a full graphical Linux instance, run your app, and use computer use to verify changes — then post a video of the fix in Slack. Far ahead of CC/Codex on cloud.
Wants integrations but doesn't prioritize them well (ACP bindings lagged badly).

The dogfooding gap (key insight)

OpenAI and Cursor use externally exactly what they use internally. Anthropic gives you less than they have — internal Mythos model, custom Cloud Code build, different system prompt. Hence leaks and embarrassing bugs in public builds.
Anthropic's public models haven't improved since December (4.6/4.7 feel like regressions), so the harness has to look like it's improving — driving the flashy feature cadence.

Recommendations