Claude Opus 5.5評測:反編譯老遊戲,GTA秒變瀏覽器版 | Claude Opus 5.5 Review: Decompiling GTA, Halo for Browser
By Kit 小克 | AI Tool Observer | 2026-10-19
🇹🇼 Claude Opus 5.5評測:反編譯老遊戲,GTA秒變瀏覽器版
這兩週整個開發者社群都在瘋傳同一件事:Claude Opus 5.5被拿來反編譯老遊戲,讓原本要花上好幾個月的逆向工程工作,變成「丟著跑一整晚」就能搞定的事。《GTA:罪惡都市》、《最後一戰:戰鬥進化》、《Skate 3》等PS2、PS3時代的經典作品,陸續被改寫成能直接在瀏覽器分頁裡執行的WebAssembly版本,不用下載、不用模擬器,開一個網址就能玩。
這次到底發生了什麼事
根據Kotaku等媒體報導,過去兩週內冒出了數百個由AI協助完成的遊戲反編譯/移植專案,多半集中在PS2、PS3世代的主機遊戲。做法大致是:把原始遊戲的執行檔丟給Claude Opus 5.5,讓它自己讀機器碼、還原成可讀的C/C++原始碼,再重新編譯成能跑在瀏覽器裡的WebAssembly。整個流程最花時間的反編譯階段,過去得靠工程師手動盯著組語一行一行拆,現在變成「晚上丟給AI、早上收成果」。
- GTA:罪惡都市(Grand Theft Auto: Vice City)
- 最後一戰:戰鬥進化(Halo: Combat Evolved)
- Skate 3
- 決勝時刻:黑色行動(Call of Duty: Black Ops)
老實說:「數百款」這個數字有點灌水
這是本文想提醒的重點。社群瘋傳的「數百款作品」是集體聲量堆出來的印象,不是每一個專案都經過獨立驗證,有些移植版本其實還有畫面錯位、音效不同步、特定關卡卡死等問題,距離「完整可玩」還有一段距離。開發者momo5502在自己的部落格詳細記錄了用AI代理反編譯一款射擊遊戲的過程,光是這一個專案就燒掉超過5000億個token——這提醒我們,用Claude Opus 5.5反編譯聽起來很神,但背後的運算成本跟人力除錯其實都還在,只是換了形式。
著作權灰色地帶,不能不提
反編譯本身在多數地區屬於灰色地帶,拿反編譯結果公開發布可執行的遊戲版本,更直接踩到著作權邊界,跟用什麼工具做出來的無關。這不是「AI做的所以沒事」,法律風險該怎麼算還是怎麼算。
對開發者真正有意義的地方
比起遊戲保存這件事本身,更值得關注的是:這證明了長時間、無人監督的AI代理(agentic coding)已經能在「枯燥、高容錯要求」的逆向工程任務上撐過一整晚還保持產出品質。這對老系統維護、安全研究、遺留程式碼除錯等場景,是比遊戲移植更實際的應用方向。
好不好用,試了才知道。
🇺🇸 Claude Opus 5.5 Review: Decompiling GTA, Halo for Browser
For the past two weeks, developers have been passing around the same story: Claude Opus 5.5 is being used to decompile old console games, turning a job that used to take months of manual reverse engineering into something you leave running overnight. Titles from the PS2 and PS3 era — including Grand Theft Auto: Vice City, Halo: Combat Evolved, Skate 3, and Call of Duty: Black Ops — have been rebuilt to run directly inside a browser tab, no emulator or download required.
What actually happened
According to reporting from Kotaku and others, hundreds of AI-assisted decompilation and porting projects have surfaced over the past two weeks, most targeting PS2/PS3-generation titles. The workflow: feed the original game binary to Claude Opus 5.5, let it read the machine code and reconstruct readable C/C++ source, then recompile that source to WebAssembly so it runs locally in the browser. The slowest part of the pipeline — line-by-line disassembly engineers used to do by hand — now gets handed to an agent that runs unattended overnight.
- Grand Theft Auto: Vice City
- Halo: Combat Evolved
- Skate 3
- Call of Duty: Black Ops
The honest take: "hundreds of games" is doing a lot of lifting
This is the part worth slowing down on. The "hundreds of games" figure is a social-media aggregate, not an independently verified count — not every project has been confirmed to actually work end-to-end. Some ports reportedly still have broken textures, desynced audio, or levels that crash. Developer momo5502 documented decompiling a single first-person shooter with an AI agent in detail, and that one project alone burned through more than 500 billion tokens. Using Claude Opus 5.5 to decompile a game overnight sounds magical, but the compute cost and debugging effort did not disappear — they just moved off a human clock and onto a GPU bill.
The copyright problem nobody is skipping
Decompiling a binary sits in a legal gray zone in most jurisdictions, and publishing a playable executable built from that decompilation moves closer to a copyright violation — regardless of which tool did the work. "An AI did it" does not change how the law reads.
Why this actually matters
The real signal is not game preservation — it is that unattended, long-horizon agentic coding can now hold up overnight on tedious, low-error-tolerance reverse-engineering work. That is a more practical takeaway for legacy system maintenance, security research, and debugging undocumented code than for retro gaming.
好不好用,試了才知道。
Sources / 資料來源
- Kotaku: Vibe-Coded Browser Ports Of Games Work Perfectly, Unfortunately
- 500+ Billion Tokens Later: Letting AI Agents Decompile A First-Person Shooter
- Claude Opus 5.5 Game Ports Put Halo and GTA in Browser Tabs, but Verification Lags Behind
延伸閱讀 / Related Articles
- OpenAI數學論文評測:722篇齊發,3篇撤稿收場 | OpenAI Math Papers Review: 722 Dropped, 3 Retracted
- OpenCode資安事故評測:API金鑰外洩,1.4%用戶遭盜刷 | OpenCode Breach Review: API Keys Leaked, 1.4% Users Hit
- Replit預設模型評測:棄用Claude,改押OpenAI | Replit Default Model Review: Drops Claude for OpenAI
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言