OpenAI’s agent Codex, this time directly goes hard against Claude Cowork.
Codex is OpenAI’s flagship code generation model, supporting products like GitHub Copilot, and has become an indispensable AI assistant for developers worldwide. This OpenAI Codex update is very heavy.
YouTube creator Mike Russell released an actual test video, the effect is explosive.
He handed his Mac completely over to OpenAI’s newly upgraded Codex, letting GPT-5.5 control Adobe Audition to repair audio, use Photoshop to make a cover, and then use Adobe Firefly to generate an AI video.
From beginning to end, zero human operation throughout the whole process.
This is not a Demo, not a PPT, it is a real creator completely handing over his productivity toolchain to AI to run once.
OpenAI co-founder, President Greg Brockman directly shouted: “Codex is usable by everyone, computer tasks can all be done!”
Yes, a tool for writing code, suddenly wants to grab everyone’s keyboard.
AI big V Guizang stated, in one afternoon, with one sentence, Codex helped him develop a complete game.
What makes people most surprised is the way Codex processes materials: he provided a material package containing thousands of pictures, without explaining the screening method.
Codex, however, automatically integrated the pictures in each folder into an overview image, attached with file names.
This way, just looking at one picture you can grasp all material styles, and after selecting, directly call the file. This operation is really shocking, making him exclaim directly Codex is too awesome!
Netizens exclaimed directly, Codex has finally ushered in its own “Claude Code highlight moment” — a complex complete Mac application, integrating camera, microphone, screen recording, it got it done in one go.
Netizens who have used Codex simply cannot stop!
Codex Has Changed: From Code Assistant to Computer Butler
In short, in the past everyone’s understanding of Codex was very clear — just a tool for writing code. It can help you complete functions, debug bugs, generate scripts, it is the programmer’s copilot.
This upgrade directly blew the boundaries open.
The most core sentence in OpenAI’s official announcement: Codex now supports Slack integration and Google Workspace family bucket integration. Translated into human language is — it not only can write code, but also read your emails, reply to your Slack messages, operate your Google Docs and Sheets.
This sentence lets OpenAI’s ambition not be hidden anymore: it no longer positions Codex as a developer tool, but — a general computer control agent.
Just yesterday, Codex suddenly officially announced a large wave of updates.
It can cross Slack, Gmail, Calendar to automatically summarize changes, do data analysis, assist decision-making.
Can organize research materials, make spreadsheets and presentations.
Can analyze data exports, mark changed content, draft interpretation reports.
Can also compare multiple choices based on standards, track trade-offs.
OpenAI co-founder Greg Brockman, this top hacker who is used to 20 years of black screen command line terminals and sees code as his life, publicly announced: I completely fell in love with the Codex App, it has already replaced the terminal I used for 20 years.
Developers who understand, all understand, what kind of weight this is.
Such a powerful OpenAI Codex update made Altman directly post exclaiming: “Codex is experiencing a ChatGPT moment!”
Following this large wave of updates yesterday, early this morning, OpenAI Codex core member Tibo posted on X saying “Feeling codexy today”, indicating that Codex will usher in an epic update again.
Once this post came out, the programmer circle instantly boiled over!
Sure enough, not long after, OpenAI started releasing new cases again.
Using Codex to handle daily work has never been this relaxed. You can choose your role, connect the apps you use every day, and try the recommended prompt words.
Whether it is research and planning, or documents, presentations, spreadsheets, etc., Codex can provide help.
Recommend Plugins and Connect Applications Naturally
Codex will recommend useful plugins based on your role, and guide you to connect various applications, like SlackHQ, GoogleWorkspace, Microsoft365, and so on.
It is like your personal assistant, can summarize data from different applications and documents, plan the next step, draft work, organize research, or create project plans.
You can see what is happening at a glance, including task progress, files and tools used, and what to do next.
From draft to final draft, you can review in Codex as the content gradually takes shape. Open the file, propose modification suggestions, and continuously optimize and adjust in the same conversation thread.
Developer big V stated, Codex and Claude Code are very different.
If the quota is about to end, then you can execute a long-time task, even if the quota has already ended, Codex will continue to execute this task until the task is completed.
This post was directly forwarded by Altman.
Tibo also stated, between good user experience and optimizing profit margins, OpenAI chose the former.
Even, OpenAI specially released an official blog guide, introducing how to use Codex in daily work.
Claude Code’s Number One Fan Turns to Codex, Altman Applauds
Just on the same day as the Codex upgrade, another good show opened.
On X, a user spoke out her true feelings: Claude Code’s generation quality has noticeably declined in the recent three weeks, accuracy plummeted, so she is using Codex 90% of the time, feeling very satisfied.
Altman quickly appeared, responding with a Star Wars meme: “Welcome to the light side!”
Sure enough, more developers stood up to express, really don’t like using Claude, because it is very clumsy, the user interface is also always wrong, bugs are also many.
This time, developers voted with their own feet.
Codex Actual Test is Too Crazy!
Codex App developer Andrew Ambrosino bluntly said: “Codex gets everything done!”
This update, Codex automatically adapts dynamic UI for the current task, better experience:
- Slides and tables experience better
- Supports direct annotation in browsers, artifacts, and code
- Simpler to get started
- Overall design is more concise
- Performance fully improved
In the Codex in-app browser, a device toolbar was also added, making building and testing responsive apps more convenient——
The speed of browser use (about 30% improvement in subjective tests).
However, “everyone saying good is really good”, the entire internet’s first wave of actual tests has already arrived. Let us take a look first!
Taking Over the Entire Mac, Humans Entirely 0 Operation Watching
Mike Russell’s actual test is the most direct proof of this upgrade.
He gave Codex three tasks:
Task one: Audio repair. A piece of recording has obvious background noise and sibilance problems. Codex automatically opens Adobe Audition, identifies noise characteristics, applies noise reduction filtering, adjusts EQ parameters, exports the finished product.
Russell listened back afterwards and evaluated: “Professional-level repair, cleaner than my manual tuning.”
Task two: Podcast cover design. Codex opens Photoshop, automatically chooses color schemes based on the podcast theme, typesets title text, adjusts layer blending modes, outputs a cover image that can be directly uploaded.
Task three: AI video generation. Codex calls Adobe Firefly, generates video material clips based on text descriptions, automatically splices, adds transitions.
Three tasks, across three Adobe professional softwares, fully automatically completed.
Russell repeatedly emphasized a detail in the video: he didn’t touch the mouse the whole process, didn’t touch the keyboard, didn’t even switch windows. Codex itself completed all software switching and coordination at the operating system level.
“This is not AI helping me work,” Russell said, “This is AI working for me.”
What Codex hits this time is not programmers, it is everyone who relies on computers to work.
When AI can control your entire computer, the skill of “knowing how to use software” itself is depreciating in value.
Of course, Russell’s actual test is not perfect.
The video material generated by Firefly had obvious screen shaking in a few frames, Codex did not automatically identify and correct it. The text typesetting of the Photoshop cover had a problem of inconsistent font sizes during the first attempt, Codex discovered it itself and made a second adjustment to pass.
Russell’s summary is very practical: “It is not 100 points, about 85 to 90 points. But the problem is — to reach this level it took 8 minutes, doing it myself takes 2 hours.”
85 points multiplied by 8 minutes, and 100 points multiplied by 2 hours. In most scenarios, the former wins.
Codex Helps You 0 Cost Unlimited Times Shooting
Netizen Matthew Berman directly introduces how to use Codex for unlimited times shooting products, one network link can be transformed into complete e-commerce photos:
Before: A set of e-commerce product pictures costs 5,000 – 25,000 US dollars, takes 4 weeks.
Now: Input a URL, 10 minutes to produce pictures, cost is 0.
He packaged the entire system into a “Brand Shoot Kit”.
How does it turn a web link into a whole set of e-commerce photography library?
Just needs the following 7 Agent skills:
Human’s Keyboard, Finally Going to be Eliminated?
In the past, comprehensively using manual methods to debug UI, is often very exhausting of mental effort.
Every time you have to check bit by bit whether AI has broken other unrelated parts, this kind of pressure is silent.
But if we can also hand over the runtime UI behavior testing to AI to do, then the burden on the human side can get a reasonable reduction.
Now, Codex has finally brought hope!
Obviously, Codex, can already use the mouse to inspect one by one whether the UI interface or behavior is normal — the whole process is completely automated.
Netizens sighed: “This feels like ‘what people have always expected AI to be able to do’ has finally arrived.” “I feel we are gradually approaching the critical point of the next major shift.”
At the end of the video, Russell said this sentence: “When AI can control your entire computer, the skill of knowing how to use software itself is depreciating in value.”
This time, what Codex hits is not programmers, after all programmers have long been used to AI writing code.
What is hit this time is everyone who relies on computers to work — those who make PPTs, write emails, edit audio, retouch pictures, make reports.
The previous logic was, humans learn to use tools, tools magnify human abilities. Now the logic has started to change: AI learns to use tools, humans just need to explain clearly what they want.
It can be said, Codex is not upgrading functions, it is redefining the matter of “using a computer” itself.
In Russell’s 45-minute actual test, everything that happened on that Mac — the mouse moving by itself, software switching by itself, audio rendering by itself — this scene will probably become the most concrete scene of 2026.
Previously humans used the mouse to call software, now AI uses API to call software.
