Skip to content
◆ Claude 4.5 Sonnet (Claude Code) · ChatGPT (GPT-5) + Cursor · Gemini 2.5 Pro Deep Think

Code Debugging + Refactoring.

Intent — Diagnose stack traces, refactor monster functions

The hottest LLM battleground. All three frontier models post strong SWE-bench numbers but real-world usage diverges hugely — Claude Code can agentically operate on a repo / GPT-5 has the deepest Cursor integration / Gemini 2.5 Pro digests an entire codebase before making architectural decisions.

01

Claude 4.5 Sonnet (Claude Code)

◆ Recommended prompt
Here's a TypeScript Next.js project. The /api/comments endpoint is throwing 500 sometimes — see attached stack trace + the route handler. Diagnose root cause, propose fix, then implement it across all relevant files. Run typecheck after.
Open in Claude (paste it)
✓ Strengths
  • **Claude Code agentic mode**：直接 read/write/run，免來回貼 code
  • 對程式碼 "intent" 理解深 — 抓得到 race condition / TOCTOU 等微妙 bug
  • Long-context refactor 連貫性最強
✗ Weaknesses
  • Claude Code 是 CLI / IDE plugin，不像 Cursor 內建編輯體驗
  • Web 搜尋整合需要 Computer Use
  • Pro 訂閱有訊息額度限制
When to use

全 stack debugging / refactor / agentic 任務（讓它自己跑 test、自己修）。

02

ChatGPT (GPT-5) + Cursor

◆ Recommended prompt
// In Cursor: select the failing function and use Cmd+K
Fix this race condition in the rate limiter — the SET-IF-NOT-EXISTS check has a TOCTOU window. Use atomic INCR with TTL instead.
Open in ChatGPT (paste it)
✓ Strengths
  • **Cursor / Copilot 整合最深** — IDE 內 inline 完成、ghost text、CMD+K
  • GPT-5 在 LiveCodeBench / SWE-bench 持續領先
  • o1 / o1 Pro 的深度思考對複雜邏輯題特別強
✗ Weaknesses
  • Long-context refactor 不如 Claude 連貫
  • Code Interpreter / Canvas 有時 over-engineer
  • Pro plan 才有 o1 Pro
When to use

IDE 內 daily coding / inline 補全 / 短任務：bug fix / 加 feature flag / 寫測試。

03

Gemini 2.5 Pro Deep Think

◆ Recommended prompt
[Attach: full 50k-line codebase via Repo extension]

This Next.js app is migrating from Pages Router to App Router. Read the entire codebase, identify all API routes / page components / shared utilities, and produce: (1) migration order with dependency graph (2) breaking changes list (3) per-file refactor checklist. Don't skip any file.
Open in Gemini (paste it)
✓ Strengths
  • **1M token context**：可看完整 50k 行 codebase 再給 architectural advice
  • 對 monorepo / cross-package 依賴的 understanding 比另兩個強
  • Deep Think 模式對 "先想再答" 任務有幫助
✗ Weaknesses
  • Inline IDE 整合弱於 GPT + Cursor
  • Coding agentic 工具鏈不如 Claude Code 成熟
  • 回應較慢（Deep Think 模式）
When to use

大型 refactor / migration / architecture review / 看完整 codebase 後做計畫的工作。

◆ Bottom line

Agentic repo operations → Claude Code. Cursor / Copilot inline completion → GPT-5. Refactoring a large codebase after full understanding → Gemini 2.5 Pro.

◆ More comparisons