今天，我們正式推出全新版本的 ChatGPT 圖像，由最新的旗艦級圖像生成模型驅動。無論你是由零開始創作，還是需要細緻修圖，都能更精準地獲得你所想像的結果。新模型在進行精細編輯時，能保留人物外貌等關鍵細節，同時圖像生成速度提升多達 4 倍。同時，我們亦在 ChatGPT 中推出全新「圖像」專區，令圖像創作更充滿趣味，激發靈感並令創意探索更輕鬆。
全新的圖像模型及功能現已於 ChatGPT 向所有用戶推出，並於 API 中以 gpt-image-1.5 提供。
現在，當你想編輯已上載的圖片時，模型能更可靠地遵循你的指示，包括最細微的調整。你可以只修改指定的部分，同時保持光線、構圖和人物外觀在不同輸入、輸出及之後的編輯仍然一致。
這讓你可以得到符合預期的成品，包括：更實用的相片編輯、更逼真的服裝試穿和髮型效果，以至保留原圖精髓的風格濾鏡與概念轉換。整體而言，這些改進讓 ChatGPT 成為你的「隨身創意工作室」，既能處理實際編輯需要，亦能進行富創意的重新演繹。
模型在多種修圖操作上表現出色，包括：新增和移除內容、組合、混合創作和轉換位置，能夠幫你完成所需變更，同時保留圖片原有的特色。
Combine the two men and the dog in a 2000s film camera-style photo of them looking bored at a kids birthday party.
Add chaotic kids in the background throwing things and screaming.
Change the man on the left to a hand-drawn retro anime style, the dog to plushie style, keep the man on the right and background scenery the way they are.
Put them all in OpenAI sweaters that look like this.
Now remove the two men, just keep the dog, and put them in an OpenAI livestream that looks like the attached image.
模型在創意轉換方面同樣出色，可加入或改變元素（例如文字與版面設計），在保留重要細節的同時，將構思具體呈現。無論是簡單或更複雜的概念，都可以使用這些轉換功能，你更可透過全新 ChatGPT 圖像（在新視窗中開啟）功能內的預設風格與構思輕鬆嘗試不同效果，毋須自行撰寫提示詞。
Make an old school golden age hollywood movie poster of a movie called 'codex' from the image of these two men. feel free to change their costumes to fit the times
Change the names of the actors to Wojciech Zaremba (left) and Greg Brockman (right)
Directed by Sam Altman, produced by Fidji Simo. A Feel the AGI Pictures Production.
相比初代版本，新模型在遵從指示方面更穩定，讓你能進行更精準的編輯，亦能創作更複雜的原創構圖，並確保各元素之間的關係符合原意。
新模型
draw a 6x6 grid
Make a 6 (columns) by 6 (rows) grid grid of:
Row 1: the Greek letter beta, a beach ball, a lemon, a robot, a fish tank, a frog
Row 2: a praying mantis, an expensive watch, a baththub, a pair of sunglasses, a colorful butterfly, an envelope
Row 3: a stamp, a picture frame, a steaming dumpling, the word "miracle", a pair of skis, the letter Z
Row 4: a toilet, a subway token, a mute icon, a bottle of perfume, a dragonfly, a skateboard helmet
Row 5: a Bluetooth icon, the number 13, a green heart, a rubik's cube, a Canada goose, a soldier's helmet
Row 6: a white dog, a life jacket, a knot, a keyboard, a tissue box, the number 14
之前模型
draw a 6x6 grid
Make a 6 (columns) by 6 (rows) grid grid of:
Row 1: the Greek letter beta, a beach ball, a lemon, a robot, a fish tank, a frog
Row 2: a praying mantis, an expensive watch, a baththub, a pair of sunglasses, a colorful butterfly, an envelope
Row 3: a stamp, a picture frame, a steaming dumpling, the word "miracle", a pair of skis, the letter Z
Row 4: a toilet, a subway token, a mute icon, a bottle of perfume, a dragonfly, a skateboard helmet
Row 5: a Bluetooth icon, the number 13, a green heart, a rubik's cube, a Canada goose, a soldier's helmet
Row 6: a white dog, a life jacket, a knot, a keyboard, a tissue box, the number 14
模型的文字渲染能力再進一步提升，能更清晰地處理密集或細小文字。
There is a newspaper on a desk. The newspaper shows the markdown below laid out as a **natural** newspaper article. Preserve all content, formatting, and numbers exactly. The image should be tall.
# Introducing GPT‑5.2
### *The most advanced frontier model for professional work and long-running agents*
**December 11, 2025**
---
We are introducing **GPT‑5.2**, the most capable model series yet for professional knowledge work.
Already, the average ChatGPT Enterprise user says AI saves them 40–60 minutes a day, and heavy users say it saves them more than 10 hours a week. We designed GPT‑5.2 to unlock even more economic value for people; it’s better at creating spreadsheets, building presentations, writing code, perceiving images, understanding long contexts, using tools, and handling complex, multi-step projects.
GPT‑5.2 sets a new state of the art across many benchmarks, including GDPval, where it outperforms industry professionals at well-specified knowledge work tasks spanning 44 occupations.
---
## Benchmark highlights
| Benchmark | Domain | GPT‑5.2 Thinking | GPT‑5.1 Thinking |
|---|---|---:|---:|
| GDPval (wins or ties) | Knowledge work tasks | **70.9%** | 38.8% (GPT‑5) |
| SWE-Bench Pro (public) | Software engineering | **55.6%** | 50.8% |
| SWE-bench Verified | Software engineering | **80.0%** | 76.3% |
| GPQA Diamond (no tools) | Science questions | **92.4%** | 88.1% |
| CharXiv Reasoning (w/ Python) | Scientific figure questions | **88.7%** | 80.3% |
| AIME 2025 (no tools) | Competition math | **100.0%** | 94.0% |
| FrontierMath (Tier 1–3) | Advanced mathematics | **40.3%** | 31.0% |
| FrontierMath (Tier 4) | Advanced mathematics | **14.6%** | 12.5% |
| ARC-AGI-1 (Verified) | Abstract reasoning | **86.2%** | 72.8% |
| ARC-AGI-2 (Verified) | Abstract reasoning | **52.9%** | 17.6% |
---
Notion, Box, Shopify, Harvey, and Zoom observed that GPT‑5.2 demonstrates state-of-the-art long-horizon reasoning and tool-calling performance. Databricks, Hex, and Triple Whale found GPT‑5.2 to be exceptional at agentic data science and document analysis tasks. Cognition, Warp, Charlie Labs, JetBrains, and Augment Code report that GPT‑5.2 delivers state-of-the-art agentic coding performance, with measurable improvements in areas such as interactive coding, code reviews, and bug finding.
In ChatGPT, GPT‑5.2 Instant, Thinking, and Pro will begin rolling out today, starting with paid plans. In the API, they are available now to all developers.
Overall, GPT‑5.2 brings significant improvements in general intelligence, long-context understanding, agentic tool-calling, and vision—making it better at executing complex, real-world tasks end-to-end than any previous model.
Now change the article to the markdown below:
# Introducing GPT‑Image‑1.5
### *The new and improved ChatGPT Images*
**December 16, 2025**
---
Today, we’re introducing a new and improved version of ChatGPT Images, powered by our best image generation model yet. With stronger instruction following and more precise editing, ChatGPT Images delivers the changes you ask for while keeping important details like facial likeness consistent across edits—now with generation speeds up to **4× faster**, making it easier to iterate and explore ideas with less waiting.
This is our most capable general-purpose text-to-image model to date, with more expressive transformations, improved dense text rendering, and more natural-looking results. Whether you’re making a tiny fix or a total reinvention, you can simply say what you want—or choose from preset styles and ideas in the new Images experience—and ChatGPT handles the rest, delivering results that are both useful and compelling, and better match your intent.
The new Images model and experience is beginning to roll out today in ChatGPT for all users, and in the API as **GPT‑Image‑1.5**.
---
## Results that match your intent
The model now follows instructions more reliably—down to the small details—changing what you ask for while able to keep elements like lighting, composition, and likeness consistent across inputs, outputs, and subsequent edits.
This unlocks results that match your intent—more useful photo edits, more believable clothing and hairstyle try-ons, alongside stylistic filters and conceptual transformations that retain the essence of the original image. Together, these improvements mean ChatGPT can act as a creative studio in your pocket, capable of both practical edits and expressive reimaginings.
### Editing
The model excels at different types of editing so you get the changes you want without losing what makes the image special.
### Creative Transformations
The model’s creativity shines with creative transformations, changing and adding elements—like text and layout—that help the concept come to life while maintaining important details.
### Instruction Following
The model is able to better follow instructions versus GPT Image 1.0.
### Text Rendering
The model takes another step ahead in text rendering, capable of handling denser and smaller text.
---
## A new creation space
In addition to asking for images through ChatGPT by describing what you’d like to see, we’re also introducing a dedicated Images experience in the ChatGPT sidebar to make exploring and trying images easier and quicker. This includes preset filters and trending prompts to jump-start inspiration, as well as a one-time likeness upload so you can reuse your appearance across future creations without the need to go through your camera roll again.
Together, these upgrades let you create images that better match your vision, from small edits to full reimaginings. Images now render up to four times faster, and you can continue generating new images while others are still in progress—so you can explore more ideas without waiting.
模型亦在其他方面有所提升，令輸出結果更即時可用，例如能更好地處理大量細小人臉，以及整體畫面看起來更自然。
新模型
make a scene in chelsea, london in the 1970s, photorealistic, everything in focus, with tons of people, and a bus with an advertisement for "ImageGen 1.5" with the OpenAI logo and subtitle "Create what you imagine". Hyper-realistic amateur photography, iPhone snapshot quality…
之前模型
make a scene in chelsea, london in the 1970s, photorealistic, everything in focus, with tons of people, and a bus with an advertisement for "ImageGen 1.5" with the OpenAI logo and subtitle "Create what you imagine". Hyper-realistic amateur photography, iPhone snapshot quality…
除了透過在訊息中描述你想看到的內容來生成圖片外，我們亦在 ChatGPT 中推出一個專屬的圖像（在新視窗中開啟）專區，透過流動應用程式側邊欄或者 chatgpt.com 即可方便使用，令探索和嘗試不同圖片效果更快更輕鬆。當中包括數十款預設濾鏡及提示構思，助你快速啟發靈感，並會定期更新，以緊貼最新趨勢。
這些升級讓你無論是進行細微修改，還是全面重新構想，都能創作出更貼近你構思的圖像。
此模型透過更快速的圖像生成、精準的編輯，以及在多次修改中保持視覺細節一致，協助企業簡化工作流程。團隊可用來探索構思、進行針對性修改，並將複雜或較抽象的概念視覺化，支援市場推廣、設計、電子商務及內部溝通等多個應用場景。
我們重新測試了初次推出圖像生成功能時的多個範例，以評估整體表現。模型在多個情況下均展現出明顯的提升，但結果仍未臻完善。雖然這次推出標誌着重要進展，但在未來版本中仍有相當大的改進空間。
新模型
create a poster of deep sea creatures at different depths, with a vertical ocean cutaway, styled in a beautiful japanese detailed anime style
之前模型
create a poster of deep sea creatures at different depths, with a vertical ocean cutaway, styled in a beautiful japanese detailed anime style
仍然有部分科學上的不準確之處，但整體正確率約為 70%，圖像表現更加生動，並能避免過早裁剪的情況。
API 中的 gpt-image-1.5 帶來與 ChatGPT 圖像功能相同的各項改進，在圖像保留及編輯能力方面均較 GPT Image 1 更為出色。
你會發現在多次編輯中，能更一致地保留品牌標誌及關鍵視覺元素，因此特別適合用於圖像設計和標誌製作等市場推廣和品牌相關工作；同時亦非常適合電子商貿團隊，從單一來源圖片生成完整的產品圖像目錄（不同款式、場景及拍攝角度）。
與 GPT Image 1 相比，GPT Image 1.5 的圖像輸入及輸出成本現已降低 20%，讓你在相同預算下可生成及反覆調整更多圖片。你可於 OpenAI Playground（在新視窗中開啟） 試用這個全新模型，或閱讀提示詞指南（在新視窗中開啟）以獲取靈感。
來自創意工具、電子商貿、市場推廣軟件等不同行業的企業及初創公司，都已開始使用 GPT Image 1.5。以下分享其中一些例子。
新模型
之前模型
「GPT 圖像 1.5 能生成高逼真度的圖像，並嚴格遵從提示詞要求，同時保留構圖、光影及細緻入微的細節。生成效果清晰、真實且穩定可靠，有助在 Wix 等平台上加快由構思到實際製作的工作流程。根據我們的測試結果，以及 Wix 目前的主要使用場景來看，其一致性與整體質素表現出眾，足以媲美現時市面上的旗艦級圖像生成模型。」
Wix 人工智能研究與數據科學主管 Hila Gat
全新的 ChatGPT 圖像功能現正於全球陸續向所有 ChatGPT 及 API 用戶推出，並已在各個平台全面登場。此功能可跨多個模型運作，毋須額外選擇任何設定即可使用。
我們相信，圖像生成所能帶來的可能性仍處於起步階段。今次更新是重要的一步，未來還會陸續推出更多功能，包括更細緻的編輯，以及在不同語言下呈現更豐富、更高細節的輸出。
作者
Contributors
Project Leadership
Gabriel Goh — Research Lead
Adele Li — Product Lead
Bill Peebles — Sora Lead
Aditya Ramesh — World Simulation Lead
Mark Chen — Chief Research Officer
Prafulla Dhariwal — Multimodal Lead
Core Team
Alex Fang, Alex Yu, Ben Wang, Bing Liang, Boyuan Chen, Charlie Nash, David Medina, Dibya Bhattacharjee, Jianfeng Wang, Kenji Hata, Kiwhan Song, Mengchao Zhong, Mike Starr, Yuguang Yang
Research Contributors
Bram Wallace, Dmytro Okhonko, Haitang Hu, Kshitij Gupta, Li Jing, Lu Liu, Peter Zhokhov, Qiming Yuan, Senthil Purushwalkam, Yizhen Zhang
Core Inference
Adam Tart, Alyssa Huang, Andrew Braunstein, Jane Park, Karen Li, Tomer Kaftan
Research Collaborators
Aditya Ramesh, Alex Nichol, Andrew Kondrich, Andrew Liu, Benedikt Winter, Bill Peebles, Connor Holmes, Cyril Zhang, Daniel Geng, Eric Mintun, James Betker, Jamie Kiros, Manuka Stratta, Martin Li, Raoul de Liedekerke, Ricky Wang, Ruslan Vasilev, Vladimir Chalyshev, Welton Wang, Wyatt Thompson, Yaming Lin
Inference Collaborators
Jiayu Bai, Kevin King, Stanley Hsieh, Weiyi Zheng
Data & Evaluation
Alexandra Barr, Aparna Dutta, Arshi Bhatnagar, Chao Yu, Charlotte Cole, Dragos Oprica, Emma Tang, Gowrishankar Sunder, Henry Baer, Ian Sohl, James Park Lennon, Jason Xu, Peilin Yang, Somay Jain, Szi-chieh Yu, Wesam Manassra, Xiaolei Zhu, Yilei Qian
Applied
Affonso Reis, Alan Gou, Alexandra Vodopianova, Amandeep Grewal, Andi Liu, Andrew Sima, Angus Fletcher, Antonia Woodford, Arun Eswara, Benny Wong, Bharat Rangan, Boyang Niu, Bridget Collins, Bryan Brandow, Callie Riggins Zetino, Chris Wendel, Ethan Chang, Gilman Tolle, Greg Hochmuth, Ibrahim Okuyucu, Jesse Chand, Jesse Hendrickson, Jiayu Bai, Jimmy Lin, Johan Cervantes, Kan Wu, Liam Esparraguera, Maja Wichrowska, Matthew Ferrari, Murat Yesildal, Nikunj Handa, Nithanth Kudige, Ola Okelola, Osman Khwaja, Peter Argany, Peter Bakkum, Peter Vidani, Richard Zadorozny, Rohan Sahai, Savelii Bondini, Sean Chang, Vickie Duong, Victoria Huang, Xiaolin Hao, Xueqing Li
Safety, Safety Systems, Integrity, Policy & Trust
Abby Fanlo Susk, Adam Wells, Aleah Houze, Annie Cheng, Artyi Xu, Carolina Paz, David Abelman, Femi Alamu, Jay Wang, Jeremiah Currier, Jesika Haria, Mariya Guryeva, Max Burkhardt, Paige Walker, Pedro Aguilar, Rutsu Koshimizu, Sam Toizer, Savannah Heon, Tom Rubin, Tonia Osadebe, Willow Primack, Zoe Stoll
Product Operations, Program Management and Governance
Antonio Di Francesco, Filippo Raso, Grace Wu, Josh Metherd, Ruth Costigan
Legal
Ally Bennett, Tony Song, Tyce Walters
Communications, Marketing, Community, Design & Creative
Akash Iyer, Alex Baker-Whitcomb, Angie Luo, Anne Oburgh, Antonia Richmond, Annie Tsang, Ashley Tyra, Bailey Richardson, Brandon McGraw, Cary Hudson, Dana Palmie, Evan Corrigan, Gaby Raila, Indgila Samad Ali, James Anderson, Jeremy Schwartz, Jordan Liss, Juan Garza, Julie Steele, Kara Zichittella, Karn Piluntanadilok, Kendal Peirce, Kim Baschet, Leah Anise, Livvy Pierce, Maria Clara M. Fleury Osorio, Minnia Feng, Nick Ciffone, Nick Forland, Niko Felix, Paige Ford, Rachel Puckett, Rishabh Aggarwal, Rusty Rupprecht, Souki Mansoor, Tasia Potasinski, Taya Christianson, Vasundhara Mudgil, Whitney Ferris, Yara Khakbaz, Zach Brock, Zoë Silverman
Special Thanks
Amy Yang, Arvin Wu, Avital Oliver, Brandon McKinzie, Chak Li, Chris Lu, David Duxin, Dian Ang Yap, Gabriel Petersson, Guillaume Leclerc, Hazel Byrne, Henry Aspegren, Jennifer Luckenbill, Ji Lin, Joseph Mo, Julius Hochmuth, Liunian (Harold) Li, Long Ouyang, Mariano López, Michael Zhang, Ravi Teja Mullapudi, Suvansh Sanjeev, Varun Shetty, Wenda Zhou
Exec
Fidji Simo, Hannah Wong, Jakub Pachocki, Jason Kwon, Johannes Heidecke, Kate Rouch, Lauren Itow, Mark Chen, Mia Glaese, Nick Ryder, Nick Turley, Prafulla Dhariwal, Sam Altman, Sulman Choudhry
