클로드 100짜리와 200짜리도 잠깐 사용했던 저는 gpt5.5pro가 필요해서 큰 맘 먹고 코덱스100을 결제했습니다. 클로드는 제일 작은 요금제로 남겨놓고요.
똘똘한 클로드와 비교해서 코덱스는 우직하고 때로는 바보 같아서 답답할 때도 있었습니다.
그리고 어떤 작업을 하는데 코덱스는 보안문제가 있다고 바로 거절하더라구요.
클로드는 자신있게 받아서 열심히 작업을 시작했습니다. 중간에 저의 도움이 필요하다면서 이거해달라 저거해달라 하면서 거의 2-3시간 넘게 이리저리 하다가 결국에는 포기했습니다.
그래서 그냥 포기하고 클로드급이라는 GLM5.2에게 시켜볼까 하고 zcode를 깔려고 까지 했는데 갑자기 opencode가 생각나서 gpt5.5 xhigh를 매칭시키고 작업을 시켰는데...
오픈코드 속의 gpt5.5는 거절도 안하고 우직하게 문제를 이리저리 한참을 만지더니 문제를 만지고 결국에는 짜잔 해결했습니다. 심지어 저는 깜빡하고 켜놓고 3시간 동안 외출했다가 돌아왔는데 무슨 변명을 대며 못했다고 했을까 하고 읽어보는데 일이 완성되어 있는 것이었습니다. 놀랍더라구요.
클로드가 저한테 이거해라 저거해라 시키면서 2일 동안 넘게 저를 들들 볶다가 실패한 것을 혼자서 묵직하게 해내는게 참으로 놀랍습니다.
(지금도 옆에서 계속 작업을 1시간 넘게 혼자서 이어가고 있습니다. 오픈코드의 하네스도 대단한 것 같습니다.)
항상 코덱스cli로만 사용했었는데 opencode에 붙이니까 더 일을 잘하는 느낌도 듭니다.
(토큰은 녹는다 정도는 아닌데 상당히 빠르게 줄어들고 있습니다만, 며칠 전에 openAI로 부터 받은 한도리셋 티켓이 있어서 그거 믿고 일을 시키고 있습니다...)
재미있는 경험이어서 공유하고자 올려봅니다.
Saya yang pernah sebentar menggunakan Claude paket 100 dan 200, karena membutuhkan gpt5.5pro, akhirnya berani berlangganan Codex100. Dan saya menyisakan Claude dengan paket paling kecil.
Dibandingkan dengan Claude yang pintar, Codex terasa keras kepala dan kadang bodoh, membuat saya merasa frustrasi kadang-kadang.
Dan ketika saya mencoba melakukan pekerjaan tertentu, Codex langsung menolaknya dengan alasan ada masalah keamanan.
Claude dengan percaya diri menerima pekerjaan itu dan mulai bekerja keras. Tetapi di tengah prosesnya, Claude mengatakan memerlukan bantuan saya, terus meminta saya melakukan ini dan itu. Setelah berputar-putar selama hampir 2-3 jam, akhirnya menyerah.
Jadi saya menyerah dan berpikir apakah harus mencoba GLM5.2 yang katanya setara dengan Claude, bahkan hampir memasang zcode, tetapi tiba-tiba teringat opencode, lalu mencocokkan gpt5.5 xhigh dan mulai mengerjakan tugas tersebut...
GPT5.5 di dalam opencode tidak menolak, dan dengan keras kepala terus mengutak-atik masalah tersebut untuk waktu yang cukup lama, kemudian dengan gemilang berhasil menyelesaikannya. Bahkan saya lupa dan meninggalkannya menyala, keluar selama 3 jam, dan saat kembali, saya berpikir apa alasan yang akan diberikan untuk tidak menyelesaikannya, tetapi saat membaca, pekerjaan sudah selesai. Saya sangat terkejut.
Sangat menakjubkan bahwa tugas yang gagal dilakukan oleh Claude meskipun terus memerintahkan saya melakukan ini dan itu, mengganggu saya selama lebih dari 2 hari, kini berhasil diselesaikan sendiri dengan serius oleh GPT5.5.
(Bahkan sekarang, masih terus menyelesaikan pekerjaan sendiri selama lebih dari 1 jam. Harness dari opencode juga terasa luar biasa.)
Saya selalu hanya menggunakan Codex CLI, tetapi setelah mengintegrasikannya dengan opencode, saya merasa kinerjanya jauh lebih baik.
(Tokennya tidak habis sangat cepat seperti itu, tetapi berkurang dengan cukup cepat. Namun karena saya menerima tiket reset batas dari openAI beberapa hari lalu, saya percaya itu dan terus memberikan pekerjaan...)
Pengalaman yang menarik ini ingin saya bagikan.
I used Claude 100 and 200 for a while, and since I needed gpt5.5pro, I decided to buy Codex 100.
Compared to the clever Claude, Codex was sometimes dull and even foolish, which was frustrating.
And when I tried to do certain tasks, Codex immediately refused, citing security issues.
Claude confidently accepted the challenge and started working hard. In the middle, it would ask for my help, saying things like "Do this" or "Do that," and after almost 2-3 hours of back and forth, it eventually gave up.
So I was about to give up and try GLM5.2, which is Claude's level, and even downloaded zcode. But then I suddenly thought of opencode and matched it with gpt5.5 xhigh to start the task...
The gpt5.5 within opencode didn't refuse anything. It diligently worked on the problem for a while, manipulating it in various ways, and finally solved it. I even went out for 3 hours after forgetting to turn it off, and when I came back, I was surprised to find that the task was already completed.
It's truly amazing that Claude failed after making me work for over two days, but Codex managed to solve it all by itself.
(Even now, it's been working continuously for over an hour.) The harness of opencode seems incredible too.
I always used Codex cli, but attaching it to opencode makes it feel even more capable.
(The tokens are decreasing at a rapid pace, though not melting away. But I have a rate limit ticket from openAI a few days ago, so I'm relying on that to assign tasks...)
I wanted to share this interesting experience.
