XDA writer Joe Rice-Jones published a controlled check: one model—Claude Sonnet 5.5 at high reasoning—inside three coding harnesses, Claude Code, Codex, and Cursor. The job was a buggy Kindle / Home Assistant WebSocket proxy (plus a second prompt for status / offline UI), with about $10 and 30 minutes per tool. His own hidden test suite, kept outside folders the agents could see, scored 16/16 for every finish. This is a third-party single-run write-up.
The three tools did not reach the model the same way. Claude Code talked to Claude on its native path (third-party models would add an extra translation layer the author did not trust for a fair test). Cursor needed a subscription to use third-party models and billed Sonnet 5.5 via API credits. Codex reached the same model through OpenRouter. Claude Code’s default reasoning was medium, so he raised it to high to match Cursor and Codex.
Where the ask was small and clear, the three tools wrote nearly the same code—including shared function names the author treats as a Sonnet 5.5 fingerprint. They also added the same unrequested fixes (disconnect banner, malformed JSON, return-versus-continue, and a first-connect proxy check). Open-ended choices diverged: how much to rewrite, how much defensive code, how many self-tests. Agent-written tests: Codex 10, Cursor 33, Claude Code 42 (the same suite run three times). Wall time about 10:07 / 13:01 / 14:33. Rough spend: Codex $3.73 (1.64M tokens), Cursor $1.11 (3.31M—author suspects heavy cache reads), Claude Code about $0.96 as an API-equivalent estimate (2.55M). The original UI strings were German; Codex and Cursor kept German, Claude Code wrote English. The author prefers Claude Code for day-to-day work, but says the gap is not large enough to crown an objective winner—and still does most agentic coding in Crush. We are not picking a winner either. We have not re-run this experiment yet. If you have tried a similar comparison, you can reach the site owner through the contact link in the footer. Related reading: Claude Code vs Codex: What $20 Gets You and Which to Pick and How Claude Code and Cursor Compare on Price and Limits (October 2026).
