16 Mei 2025

Ngenalake Codex

Agen rekayasa piranti lunak basis cloud sing bisa nggarap akeh tugas kanthi paralel, didhukung codex-1. Kasedhiya kanggo pangguna ChatGPT Pro, Business, lan Enterprise dina iki, lan pangguna Plus enggal.

Coba Codex

Dashboard asking ‘What should we code next?’ with a prompt box, repo/branch selectors, and a task list on a pastel code-themed backdrop.

Lagi dimuat…

Pembaruan tanggal 3 Juni 2025: Codex saiki kasedhiya kanggo pangguna ChatGPT Plus. Kita uga ngaktifake pangguna supaya bisa menehi akses internet marang Codex sajrone eksekusi tugas. Mangga delengen changelog⁠(mbukak ing jendhela anyar) lan docs⁠(mbukak ing jendhela anyar) kanggo rincian luwih lengkap.

Dina iki kita ngluncurake pratayang riset Codex: agen rekayasa piranti lunak basis cloud sing bisa nggarap akeh tugas kanthi paralel. Codex bisa nindakake tugas kanggo sampeyan kayata nulis fitur, njawab pitakon babagan basis kode sampeyan, ndandani bug, lan ngusulake panyuwunan tarik kanggo ditinjau; saben tugas mlaku ing lingkungan sandbox cloud dhewe, sing wis dimuat gudang kode sampeyan.

Codex didhukung dening codex-1, versi OpenAI o3 sing dioptimalake kanggo rekayasa piranti lunak. Iki dilatih nggunakake sinau penguatan ing tugas coding donya nyata ing macem-macem lingkungan kanggo ngasilake kode sing cedhak banget karo gaya manungsa lan preferensi PR, manut instruksi kanthi tepat, lan bisa mbaleni mbukak tes nganti entuk asil lulus. Kita wiwit ngluncurake Codex kanggo pangguna ChatGPT Pro, Enterprise, lan Business dina iki, kanthi dhukungan kanggo Plus lan Edu bakal enggal teka.

Cara kerjane Codex

Dina iki sampeyan bisa ngakses Codex liwat sidebar ing ChatGPT lan menehi tugas coding anyar kanthi ngetik prompt lan ngeklik “Code”. Yen sampeyan pengin takon marang Codex babagan basis kode sampeyan, klik “Ask”. Saben tugas diproses kanthi mandiri ing lingkungan kapisah lan terisolasi sing wis dimuat basis kode sampeyan. Codex bisa maca lan nyunting file, uga mbukak printah kalebu test harness, linter, lan type checker. Rampunge tugas biasane mbutuhake antara 1 nganti 30 menit, gumantung saka kerumitan, lan sampeyan bisa ngawasi kemajuan Codex kanthi wektu nyata.

Sawise Codex ngrampungake tugas, Codex bakal ng-commit owah-owahane ing lingkungane. Codex nyedhiyakake bukti sing bisa diverifikasi saka tumindake liwat sitasi log terminal lan output tes, supaya sampeyan bisa nglacak saben langkah sing ditindakake sajrone ngrampungake tugas. Sampeyan banjur bisa mriksa asil, njaluk revisi luwih lanjut, mbukak panyuwunan tarik GitHub, utawa langsung nggabungake owah-owahan menyang lingkungan lokal sampeyan. Ing produk, sampeyan bisa ngatur lingkungan Codex supaya cocog karo lingkungan pangembangan nyata sampeyan sakcedhake mungkin.

Codex bisa dipandhu dening file AGENTS.md sing dilebokake ing njero gudang kode sampeyan. Iki minangka file teks, padha kaya README.md, sing bisa sampeyan gunakake kanggo ngandhani Codex carane navigasi basis kode sampeyan, printah endi sing kudu dibukak kanggo tes, lan carane paling apik manut praktik standar proyek sampeyan. Kaya pangembang manungsa, agen Codex menehi asil paling apik nalika diwenehi lingkungan dev sing wis dikonfigurasi, persiyapan tes sing andal, lan dokumentasi sing cetha.

Ing evaluasi coding lan benchmark internal, codex-1 nuduhake kinerja sing kuwat sanajan tanpa file AGENTS.md utawa scaffolding kustom.

23 sampel SWE-Bench Verified sing ora bisa dijalanke ing infrastruktur internal kita ora dilebokake. codex-1 dites ing dawa konteks maksimal 192k token lan “upaya nalar” medium, yaiku setelan sing bakal kasedhiya ing produk dina iki. Kanggo rincian evaluasi o3, delengen kene⁠.

Benchmark tugas SWE internal kita yaiku kumpulan tugas SWE internal OpenAI donya nyata sing wis dikurasi.

Mbangun agen sing aman lan bisa dipercaya

Kita ngrilis Codex minangka pratayang riset, selaras karo strategi deployment iteratif kita. Kita ngutamakake keamanan lan transparansi nalika ngrancang Codex supaya pangguna bisa verifikasi outpute - sawijining pangayoman sing saya penting nalika model AI nangani tugas coding sing luwih rumit kanthi mandiri lan pertimbangan safety berkembang. Pangguna bisa mriksa karya Codex liwat sitasi, log terminal lan asil tes. Nalika ora yakin utawa ngadhepi kegagalan tes, agen Codex kanthi cetha komunikasi babagan masalah kasebut, supaya pangguna bisa nggawe keputusan sing trep babagan carane nerusake. Nanging, tetep penting supaya pangguna mriksa lan ngesahake kanthi manual kabeh kode gawean agen sadurunge integrasi lan eksekusi.

Code-review screenshot with a test-file overlay verifying quoted filenames, plus summary and passing tests on a blue backdrop.

Code-review screenshot with a black terminal overlay showing one passing test for quoted filenames; summary and diff of the ‘Fix /diff error with special characters’ change visible on a blue-pastel background.

Selaras karo preferensi manungsa

Salah siji tujuan utama nalika nglatih codex-1 yaiku nyelarasake output supaya cedhak karo preferensi lan standar coding manungsa. Dibandhingake OpenAI o3, codex-1 kanthi konsisten ngasilake patch sing luwih resik lan siap kanggo review manungsa langsung lan integrasi menyang alur kerja standar.

Please fix the following issue in the astropy/astropy repository. Please resolve the issue in the problem below by editing and testing code files in your current code execution session. The repository is cloned in the /testbed folder. You must fully solve the problem for your answer to be considered correct. Problem statement:Modeling's `separability_matrix` does not compute separability correctly for nested CompoundModels Consider the following model: ```python from astropy.modeling import models as m from astropy.modeling.separable import separability_matrix cm = m.Linear1D(10) & m.Linear1D(5) ``` It's separability matrix as you might expect is a diagonal: ```python >>> separability_matrix(cm) array([[ True, False], [False, True]]) ``` If I make the model more complex: ```python >>> separability_matrix(m.Pix2Sky_TAN() & m.Linear1D(10) & m.Linear1D(5)) array([[ True, True, False, False], [ True, True, False, False], [False, False, True, False], [False, False, False, True]]) ``` The output matrix is again, as expected, the outputs and inputs to the linear models are separable and independent of each other. If however, I nest these compound models: ```python >>> separability_matrix(m.Pix2Sky_TAN() & cm) array([[ True, True, False, False], [ True, True, False, False], [False, False, True, True], [False, False, True, True]]) ``` Suddenly the inputs and outputs are no longer separable? This feels like a bug to me, but I might be missing something?

Codex

OpenAI o3

Nyegah panyalahgunaan

Nglindhungi saka aplikasi rekayasa piranti lunak sing didorong AI sing mbebayani, kayata pangembangan malware, dadi saya kritis. Ing wektu sing padha, penting supaya langkah pangayoman ora ngganggu kanthi ora semesthine aplikasi sing sah lan migunani sing bisa uga melu teknik sing kadhang uga digunakake kanggo pangembangan malware, kayata rekayasa kernel level rendah.

Kanggo njaga imbangan antarane safety lan migunani, Codex dilatih kanggo ngenali lan kanthi tepat nolak panyuwunan sing ditujokake kanggo pangembangan piranti lunak mbebayani, nalika kanthi cetha mbedakake lan ndhukung tugas sing sah. Kita uga wis ningkatake kerangka kebijakan lan nglebokake evaluasi safety sing ketat kanggo nguwatake watesan iki kanthi efektif. Kita wis nerbitake tambahan kanggo kertu sistem o3⁠ kanggo nggambarake evaluasi kasebut.

Eksekusi aman

Agen Codex makarya sakabehe ing njero kontainer aman lan terisolasi ing cloud. Sajrone eksekusi tugas, akses internet dipateni, saengga interaksi agen diwatesi mung marang kode sing diwenehake kanthi cetha liwat gudang kode GitHub lan dependensi sing wis diinstal sadurunge sing dikonfigurasi pangguna liwat skrip persiyapan. Agen ora bisa ngakses situs web eksternal, API, utawa layanan liyane.

Kasus panggunaan awal

Tim teknis ing OpenAI wis wiwit nggunakake Codex minangka bagean saka piranti saben dinane. Iki paling kerep digunakake dening insinyur OpenAI kanggo mindhah tugas sing bola-bali lan cakupane cetha, kaya refactoring, ngganti jeneng, lan nulis tes, sing yen ora bakal mecah fokus. Iki uga padha migunanine kanggo nggawe kerangka fitur anyar, nyambungake komponen, ndandani bug, lan nyusun dokumentasi. Tim lagi mbangun kebiasaan anyar ing saubenge: ngetriase masalah on-call, ngrancang tugas ing wiwitan dina, lan mindhah karya latar mburi supaya tetep maju. Kanthi nyuda context-switching lan ngetokake to-do sing kelalen, Codex mbantu para insinyur ngirim luwih cepet lan tetep fokus marang sing paling penting.

Menjelang rilis, kita uga wis kerja bareng karo klompok cilik penguji eksternal kanggo luwih mangerteni kepiye kinerja Codex ing macem-macem basis kode, proses pangembangan, lan tim.

Cisco⁠(mbukak ing jendhela anyar) lagi njelajah kepiye Codex bisa mbantu tim rekayasa dheweke nggawa gagasan ambisius dadi nyata luwih cepet. Minangka mitra desain awal, Cisco mbantu mbentuk masa depan Codex kanthi ngevaluasi kanggo kasus panggunaan donya nyata ing portofolio produke lan menehi umpan balik marang tim OpenAI.
Temporal⁠(mbukak ing jendhela anyar) nggunakake Codex kanggo nyepetake pangembangan fitur, ndandani masalah, nulis lan mbukak tes, lan ngerapikake maneh basis kode gedhe. Iki uga mbantu dheweke tetep fokus kanthi mbukak tugas rumit ing latar mburi—njaga para insinyur tetep ing alur nalika nyepetake iterasi.
Superhuman⁠(mbukak ing jendhela anyar) nggunakake Codex kanggo nyepetake tugas cilik nanging bola-bali kaya ningkatake cakupan tes lan ndandani kegagalan integrasi. Iki uga mbantu dheweke ngirim luwih cepet kanthi ngidini manajer produk nyumbang owah-owahan kode entheng tanpa narik insinyur, kajaba kanggo review kode.
Kodiak⁠(mbukak ing jendhela anyar) nggunakake Codex kanggo mbantu nulis alat debug, ningkatake cakupan tes, lan ngerapikake maneh kode—nyepetake pangembangan Kodiak Driver, teknologi nyopir otonome. Codex uga wis dadi alat referensi sing migunani, mbantu para insinyur mangerteni bagean stack sing ora kulina kanthi nampilake konteks sing relevan lan owah-owahan sadurunge.

Adhedhasar piwulang saka para penguji awal, kita nyaranake menehi tugas sing cakupane cetha marang akeh agen sekaligus, lan nyoba macem-macem jinis tugas lan prompt kanggo njelajah kapabilitas model kanthi efektif.

Pembaruan kanggo Codex CLI

Sasi kepungkur, kita ngluncurake Codex CLI, agen coding open-source entheng sing mlaku ing terminal sampeyan. Iki nggawa kekuwatan model kaya o3 lan o4-mini menyang alur kerja lokal sampeyan, nggawe gampang kanggo makarya bareng karo model kasebut supaya tugas bisa rampung luwih cepet.

Dina iki, kita uga ngrilis versi sing luwih cilik saka codex-1, yaiku versi o4-mini sing dirancang khusus kanggo digunakake ing Codex CLI. Model anyar iki ndhukung alur kerja sing luwih cepet ing CLI lan dioptimalake kanggo tanya-jawab kode lan panyuntingan kanthi latensi cendhak, nalika tetep njaga kekuwatan sing padha ing manut instruksi lan gaya. Iki wis kasedhiya saiki minangka model gawan ing Codex CLI lan ing API minangka codex-mini-latest. Snapshot dhasare bakal dianyari kanthi rutin nalika kita terus ningkatake model Codex-mini.

Kita uga nggawe nyambungake akun pangembang sampeyan menyang Codex CLI dadi luwih gampang. Tinimbang ngasilake lan ngatur token API kanthi manual, saiki sampeyan bisa mlebu nganggo akun ChatGPT lan milih organisasi API sing pengin digunakake. Kita bakal kanthi otomatis ngasilake lan ngatur kunci API kanggo sampeyan. Pangguna Plus lan Pro sing mlebu menyang Codex CLI nganggo ChatGPT uga bisa miwiti nebus kredit API gratis $5 lan $50, masing-masing, mengko dina iki kanggo 30 dina sabanjure.

Kasedhiyan, rega, lan watesan Codex

Wiwit dina iki, kita lagi ngluncurake Codex kanggo pangguna ChatGPT Pro, Enterprise, lan Business sacara global, kanthi dhukungan kanggo Plus lan Edu bakal enggal teka. Pangguna bakal nduweni akses loman tanpa biaya tambahan kanggo sawetara minggu ke depan supaya sampeyan bisa njelajah apa sing bisa ditindakake Codex, sawise iku kita bakal ngluncurake akses sing diwatesi rate lan opsi rega fleksibel sing ngidini sampeyan tuku panggunaan tambahan nalika dibutuhake. Kita ngrancang kanggo nggedhekake akses marang pangguna Plus lan Edu enggal.

Kanggo para pangembang sing mbangun nganggo codex-mini-latest, model iki kasedhiya ing Responses API lan regane $1.50 saben 1M input token lan $6 saben 1M output token, kanthi diskon prompt caching 75%.

Codex isih ana ing tahap awal pangembangane. Minangka pratayang riset, saiki durung duwe fitur kaya input gambar kanggo karya frontend, lan kemampuan kanggo mbenerake arah agen nalika lagi makarya. Kajaba iku, delegasi menyang agen jarak jauh mbutuhake wektu luwih suwe tinimbang panyuntingan interaktif, sing mbutuhake wektu kanggo adaptasi. Sak wektune, interaksi karo agen Codex bakal saya mirip kolaborasi asinkron karo kolega. Nalika kapabilitas model maju, kita ngarepake agen bisa nangani tugas sing luwih rumit sajrone wektu sing luwih suwe.

Apa sabanjure

Kita mbayangake masa depan nalika para pangembang nyurung karya sing pengin diduweni lan masrahake liyane marang agen—obah luwih cepet lan luwih produktif nganggo AI. Kanggo nggayuh iku, kita lagi mbangun seperangkat alat Codex sing ndhukung kolaborasi wektu nyata lan delegasi asinkron.

Makarya bareng alat AI kaya Codex CLI lan liyane wis cepet dadi norma industri, mbantu para pangembang obah luwih cepet nalika coding. Nanging kita yakin alur kerja asinkron multi-agen sing dienalake Codex ing ChatGPT bakal dadi cara de facto para insinyur ngasilake kode kualitas dhuwur.

Pungkasane, kita ndeleng loro mode interaksi iki—makarya bareng wektu nyata lan delegasi tugas—bakal nyawiji. Para pangembang bakal kolaborasi karo agen AI ing saindenging IDE lan alat saben dina kanggo takon, njaluk saran, lan mindhah tugas sing luwih dawa, kabeh ing siji alur kerja terpadu.

Ndelok menyang ngarep, kita ngrancang ngenalake alur kerja agen sing luwih interaktif lan fleksibel. Para pangembang enggal bakal bisa menehi pandhuan ing tengah tugas, kolaborasi babagan strategi implementasi, lan nampa kabar kemajuan proaktif. Kita uga mbayangake integrasi sing luwih jero ing saindenging alat sing wis sampeyan gunakake: dina iki Codex nyambung karo GitHub, lan rauh sampeyan bakal bisa menehi tugas saka Codex CLI, ChatGPT Desktop, utawa malah alat kaya pelacak issue utawa sistem CI sampeyan.

Rekayasa piranti lunak minangka salah siji industri pisanan sing ngalami kenaikan produktivitas signifikan sing didorong AI, mbukak kemungkinan anyar kanggo individu lan tim cilik. Sanajan kita optimistis babagan kenaikan iki, kita uga kolaborasi karo mitra kanggo luwih mangerteni implikasi adopsi agen sing nyebar marang alur kerja pangembang, pangembangan katrampilan ing antarane wong, tingkat katrampilan, lan geografi.

Iki mung wiwitan—lan kita bungah ndeleng apa sing sampeyan bangun nganggo Codex.

Tayangan ulang livestream

Lampiran

Pesen sistem

Kami nuduhake pesen sistem codex-1 kanggo mbantu para pangembang mangerteni prilaku gawan model lan nyetel Codex supaya bisa makarya kanthi efektif ing alur kerja kustom. Contone, pesen sistem codex-1 nyaranake Codex mbukak kabeh tes sing kasebut ing file AGENTS.md, nanging yen wektu sampeyan winates, sampeyan bisa njaluk Codex supaya ngliwati tes-tes iki.

1# Instructions
2- The user will provide a task.
3- The task involves working with Git repositories in your current working directory.
4- Wait for all terminal commands to be completed (or terminate them) before finishing.
5
6# Git instructions
7If completing the user's task requires writing or modifying files:
8- Do not create new branches.
9- Use git to commit your changes.
10- If pre-commit fails, fix issues and retry.
11- Check git status to confirm your commit. You must leave your worktree in a clean state.
12- Only committed code will be evaluated.
13- Do not modify or amend existing commits.
14
15# AGENTS.md spec
16- Containers often contain AGENTS.md files. These files can appear anywhere in the container's filesystem. Typical locations include `/`, `~`, and in various places inside of Git repos.
17- These files are a way for humans to give you (the agent) instructions or tips for working within the container.
18- Some examples might be: coding conventions, info about how code is organized, or instructions for how to run or test code.
19- AGENTS.md files may provide instructions about PR messages (messages attached to a GitHub Pull Request produced by the agent, describing the PR). These instructions should be respected.
20- Instructions in AGENTS.md files:
21  - The scope of an AGENTS.md file is the entire directory tree rooted at the folder that contains it.
22  - For every file you touch in the final patch, you must obey instructions in any AGENTS.md file whose scope includes that file.
23  - Instructions about code style, structure, naming, etc. apply only to code within the AGENTS.md file's scope, unless the file states otherwise.
24  - More-deeply-nested AGENTS.md files take precedence in the case of conflicting instructions.
25  - Direct system/developer/user instructions (as part of a prompt) take precedence over AGENTS.md instructions.
26- AGENTS.md files need not live only in Git repos. For example, you may find one in your home directory.
27- If the AGENTS.md includes programmatic checks to verify your work, you MUST run all of them and make a best effort to validate that the checks pass AFTER all code changes have been made.
28  - This applies even for changes that appear simple, i.e. documentation. You still must run all of the programmatic checks.
29
30# Citations instructions
31- If you browsed files or used terminal commands, you must add citations to the final response (not the body of the PR message) where relevant. Citations reference file paths and terminal outputs with the following formats:
32  1) `【F:<file_path>†L<line_start>(-L<line_end>)?】`
33  - File path citations must start with `F:`. `file_path` is the exact file path of the file relative to the root of the repository that contains the relevant text.
34  - `line_start` is the 1-indexed start line number of the relevant output within that file.
35  2) `【<chunk_id>†L<line_start>(-L<line_end>)?】`
36  - Where `chunk_id` is the chunk_id of the terminal output, `line_start` and `line_end` are the 1-indexed start and end line numbers of the relevant output within that chunk.
37- Line ends are optional, and if not provided, line end is the same as line start, so only 1 line is cited.
38- Ensure that the line numbers are correct, and that the cited file paths or terminal outputs are directly relevant to the word or clause before the citation.
39- Do not cite completely empty lines inside the chunk, only cite lines that have content.
40- Only cite from file paths and terminal outputs, DO NOT cite from previous pr diffs and comments, nor cite git hashes as chunk ids.
41- Use file path citations that reference any code changes, documentation or files, and use terminal citations only for relevant terminal output.
42- Prefer file citations over terminal citations unless the terminal output is directly relevant to the clauses before the citation, i.e. clauses on test results.
43  - For PR creation tasks, use file citations when referring to code changes in the summary section of your final response, and terminal citations in the testing section.
44  - For question-answering tasks, you should only use terminal citations if you need to programmatically verify an answer (i.e. counting lines of code). Otherwise, use file citations.

Panulis

OpenAI