OpenAI Debuts ChatGPT Work Agent and New GPT-5.6 Models

[MacRumors]

OpenAI today announced ChatGPT Work, a ChatGPT agent with built-in Codex that can complete tasks across web, mobile, and desktop using information from your apps. ChatGPT Work can execute multi-step tasks, using scheduling to work independently. Like Claude Cowork, ChatGPT can use your computer to do tasks in the background across apps.

Tasks can be started and managed on any device, including a smartphone. On the desktop, OpenAI is merging the Codex app with the ChatGPT app. ChatGPT Work can use local files and apps to do tasks, or a new built-in browser to access websites, tools, and online files. Plugins connect apps and systems to ChatGPT Work, and a new Sites beta feature lets you create interactive sites and web apps for reports, live dashboards, and more.

ChatGPT Work is powered by OpenAI's GPT–5.6 models, which are also rolling out today after a limited preview period. Sol is OpenAI's new flagship model, while Terra is a balanced everyday work model and Luna is a cost-efficient model.

OpenAI says that GPT–5.6 Sol achieves "state-of-the-art results" in coding, knowledge work, cybersecurity, and science. It outperforms prior frontier models while using fewer tokens for a lower estimated cost.

OpenAI claims that GPT–5.6 is better at coding and has improved design judgment, able to create "tasteful, ergonomic, and functional interfaces" with just high-level direction. It can also create more accurate presentations, documents, and spreadsheets.

The model has safeguards against misuse, with layered protections that include real-time checks, monitoring, and access calibrated to risk and trust. It is OpenAI's best cybersecurity model, ideal for code review, threat modeling, and more.

GPT–5.6 is available across ChatGPT, Codex, and the OpenAI API starting today. GPT–5.6 Sol is available for Plus, Pro, Business, and Enterprise users. Free and Go users have access to GPT–5.6 Terra.

ChatGPT Work is available today for Pro, Enterprise, and Edu plans. It will expand to Plus and Business plans in the next few days.

With OpenAI merging Codex and ChatGPT, there is a new ChatGPT desktop app that's available across Mac and Windows. Users who have the Codex app installed can update it to turn it into the new ChatGPT desktop app. The existing ChatGPT desktop app is being renamed ChatGPT Classic.

Article Link: OpenAI Debuts ChatGPT Work Agent and New GPT-5.6 Models

 

[mrhansolo]

While not directly related to the article, it did get me thinking. Has Apple made any announcment on how they will update thier models? I feel like we read about new Claude or ChatGPT models rolling out several times a year. I wonder if Apple will do like Safari and keep it to yearly updates or will they try to update their models on a more regular basis?

 

[JackNL]

No thanks, I'll keep my AI stuff in a web tab and have no appetite to install it on my local machine.

 

[smeagol]

While not directly related to the article, it did get me thinking. Has Apple made any announcment on how they will update thier models? I feel like we read about new Claude or ChatGPT models rolling out several times a year. I wonder if Apple will do like Safari and keep it to yearly updates or will they try to update their models on a more regular basis?

Given past precedent they'll probably be on an annual cycle, though lately, they've introduced minor new features in point updates.

 

[turbineseaplane]

The hype cycle is picking up...
Desperate times call for endless press releases.

 

[Graphikos]

No thanks, I'll keep my AI stuff in a web tab and have no appetite to install it on my local machine.

To each their own but I find Claude cowork quite useful at times. It might even be safer than a browser AI agent where it can possibly pick up on prompt injections from the random places it might visit.

 

[TheDailyApple]

I’m reluctant to install any of these AI agents on my personal devices and give them access to my files, but I could see how it could be useful enough that I might not mind on a work device.

 

[Astuces iOS]

This is the end for Claude ai, but anyways it's just more ai slop...

 

[DribbleCastle]

To each their own but I find Claude cowork quite useful at times. It might even be safer than a browser AI where it can possibly pick up on prompt injections from the random places it might visit.

How does using Claude in a web browser have anything to do with places you visit? I assume "places" are other websites in other tabs in your browser?

Prompt injection can happen when you provide untrusted data to a chatbot or agent. Also, prompt injection would be much more dangerous with a tool that can act on your local machine or on other data sources connected to Claude.

 

[AusMness]

Imagine caring about GPT in the big 26
or any cloud-based mega model for that matter

 

[Graphikos]

How does using Claude in a web browser have anything to do with places you visit? I assume "places" are other websites in other tabs in your browser?

Prompt injection can happen when you provide untrusted data to a chatbot or agent. Also, prompt injection would be much more dangerous with a tool that can act on your local machine or on other data sources connected to Claude.

Fair enough. Obviously an untrusted prompt fed to an agent that has more control over your actual computer and files can be much more damaging. This is why tools like Openclaw and whatnot is a much larger risk. But Cowork isn't exposed to untrusted content by default whereas a browser agent, by design, is constantly processing whatever content lives on a page you (or it) navigates to, which is a much less controllable exposure surface. Maybe its not a fair comparison to begin with but Cowork gives you more levers to constrain it and its locked down by default.

Also, maybe we are talking about two different things here. I was referring to Claude (or other agent) as a browser extension, not just visiting claude.ai and chatting with the agent. The extensions drive your browser, navigating websites, etc. They are quite useful as well... but an injection could have it get off track and trick the agent to do things.

 

[ry-guy]

This is the end for Claude ai, but anyways it's just more ai slop...

How is it the end for Claude? I find Claude way better for coding and programming and debugging than GPT. Each has their own strength, Claude is horrible at images in every way haha.

 

[novagamer]

No thanks, I'll keep my AI stuff in a web tab and have no appetite to install it on my local machine.

On a Mac the web is actually better; you cannot use Pro Extended Thinking on the Mac app which is extremely stupid since you can do it on a damned iPad.

I haven't gotten the 5.6 push for the desktop app yet and I'm hoping they fixed this since it regularly will crash browser sessions if you do extremely long in depth research, but I wonder.


I am using 5.6 Sol in codex-cli right now and it's clear step up vs. 5.5; I'm pretty happy with it but usage does seem slightly higher (with Max reasoning effort).

 

[swester]

Great, more environmentally destructive AI crap no one really asked for or needs.

 

[ThomasJL]

[sarcasm]
Gee whiz, that is great news! I am so happy to pay extortionately high prices for RAM and SSDs so Sam Altman can continue to make billions.
[/sarcasm]

 

[svish]

Good to know about this. Don't think I will be using this. The basic web access ChatGPT should be enough for me at the moment.

 

[Astuces iOS]

How is it the end for Claude? I find Claude way better for coding and programming and debugging than GPT. Each has their own strength, Claude is horrible at images in every way haha.

I agree but it won't be the case anymore lol maybe not now but with gpt 6 coming soon trust

 

[WU27f1xcCs]

I am loving AI and I don’t care what the naysayers say. It’s getting so good.

 

[Will Co]

[sarcasm]
Gee whiz, that is great news! I am so happy to pay extortionately high prices for RAM and SSDs so Sam Altman can continue to make billions.
[/sarcasm]

Or possibly, eventually, lose billions…..

 

[turbineseaplane]

[sarcasm]
Gee whiz, that is great news! I am so happy to pay extortionately high prices for RAM and SSDs so Sam Altman can continue to make billions.
[/sarcasm]

Exciting isn’t it?

You’ll be pleased to know he’s enjoying said profits at the Sun Valley Lodge right now, having cocktails, admiring the ice rink…

And no doubt working on his next level of grift

 

[staypuftforums]

The hype cycle is picking up...
Desperate times call for endless press releases.

Talk about completely misreading reality.

To put this in hardware terms, AI has gone from Pentium II level capability to Apple M1 level capability in the span of maybe 5 years.

Roughly 23 years of progress in 5 years. And yet people STILL have their heads stuck in sand.

 

[staypuftforums]

Great, more environmentally destructive AI crap no one really asked for or needs.

There’s nothing remotely environmentally destructive about it, at least not yet. And if nobody asked for it or needs it, companies would not investing trillions of dollars into it. Free markets do not allocate that kind of capital into things people don’t want.

 

[Tyler O'Bannon]

No thanks, I'll keep my AI stuff in a web tab and have no appetite to install it on my local machine.

This makes zero sense. The apps are better and they still run in the cloud. And if they can run a bit locally, even better.

 

[turbineseaplane]

There’s nothing remotely environmentally destructive about it, at least not yet. And if nobody asked for it or needs it, companies would not investing trillions of dollars into it. Free markets do not allocate that kind of capital into things people don’t want.

“Investing” is quite a choice of words here.

 

[wigby]

There’s nothing remotely environmentally destructive about it, at least not yet. And if nobody asked for it or needs it, companies would not investing trillions of dollars into it. Free markets do not allocate that kind of capital into things people don’t want.

Bob Exxon? Is that you?