Claude CoWork is now the one AI tool I would keep if I had to choose only one. That is not something I expected to say. For a long time, if you had forced me to pick a single AI tool, I would have said ChatGPT without hesitation. It was the most capable, the most versatile, and the most integrated into my daily work.
That has changed.
Today my answer is Claude CoWork.
To be clear, this is not because ChatGPT has gotten worse. The new ChatGPT 5.4 is excellent – in some areas I think it may be the best model out there (5.4 Pro is a beast!). But raw capability is no longer the only thing that matters to me. The question I now care about most is: which tool actually helps me get work done the way I want it done?
For me, right now, the answer is Claude CoWork.
What Makes Claude CoWork Different
The difference is not just intelligence. It is behaviour.
Claude CoWork runs locally with direct access to my files and applications. It can run and write code. It can fetch files. It can access files, run code, and interact with applications and a browser. Most importantly, it can keep working through a long, messy task instead of trying to wrap things up with a quick answer.
That last point matters more than it might sound. Most AI tools are impressive in short demos. They look smart. They produce clean answers. But real work is rarely one neat question followed by one neat answer. Real work involves files, folders, websites, spreadsheets, half-broken processes, and a lot of repetitive searching, checking, downloading, extracting, comparing, and formatting.
That is where Claude CoWork starts to feel fundamentally different. It does not just answer. It works. It keeps going until the task is completed or my credits are gone.
And the ecosystem around it is deeper than most people realise. You can write custom skill files that teach it specific workflows, formatting preferences, or domain knowledge. That means it gets better the more you invest in it. It is a genuinely rich topic that I plan to explore in more detail in future posts, but the short version is this: this is not a chatbot with a few extra features bolted on. It is a working environment. I know a handful of companies who are already investing in writing their own skill files with very good results.
A Practical Example
Suppose I need to gather commercial real estate reports for every major office market in Europe. That sounds like a single task, but it is actually a whole workflow. You have to decide which cities qualify, search for reports city by city, judge which ones are useful, download them (to my local folder, please!), extract the key data, and organise it into a spreadsheet.
ChatGPT could do that if you break it down into simple tasks. Claude CoWork can do that. Not perfectly, but really well. It can do enough of it, coherently enough, that it changes what is practical to attempt in a morning. It goes as far as filling the form on JLL’s website with my Gmail address so that JLL sends the report to my mailbox. It then opens my mailbox, finds the email, and downloads the report – neat!
It can also produce decent Word documents with real formatting, work competently inside Excel, and handle a range of file operations that would otherwise eat up hours of manual effort.
This is the shift that matters. We are moving from AI that gives you ideas to AI that executes workflows. And the important part is that this no longer requires technical knowledge. A few years ago, you needed someone who could code to automate this kind of process. Now you need someone who can define a task well, recognise good output, and check the result. That is a much more accessible skill.
The New ChatGPT 5.4 Is Genuinely Excellent
I do not want to understate how good the new ChatGPT is. I tested it on my usual benchmarks, including the one I care about most: uploading a complex PDF and asking for a detailed assessment, reasoning, and judgment. It performed extremely well. Probably the best I have seen from any model, to be honest.
On mathematics and other closed-form problems, I would put it ahead of Claude.
I had a moment last week that brought this home. I was sitting with co-authors discussing a paper, and we asked AI to work through the maths for us. The four of us just sat there and watched. It was doing the work faster and more accurately than any of us could have done it ourselves. Four (arguably top-tier) economists, watching a machine outdo them at maths. And most economists are pretty good at maths.
That was striking. These systems are not almost capable. They are already there on a lot of tasks.
Capability Is Not the Same as Usefulness
Why I Still Use ChatGPT Every Day
Here is the honest part: despite everything I have just said, I still use ChatGPT constantly.
The main reason is cost. The marginal cost of using ChatGPT is free. Claude does not feel free. And that changes behaviour more than you might expect. When something feels free, you use it for everything – quick questions, rough drafts, idle curiosity. When it feels like it costs money, you start rationing. You think twice. You save it for the tasks that really justify it. That friction is real, and it matters.
The second reason is integration. ChatGPT connects to my OneDrive, which is genuinely useful in daily work. Sometimes the best AI is not the smartest one. It is the one that can reach the document you need without friction.
So I do not think this is a winner-takes-all story. It is a maturing market where different tools are becoming good at different things, and the smart approach is to use more than one.
The Smartest Model Is Not Always the Best Tool
The smartest model I currently have access to might actually be Gemini Pro or ChatGPT Pro. There are moments when those systems feel astonishingly intelligent.
But intelligence is not enough. The questions that matter now are practical. Does it understand what I mean? Can it access the right files? Can it work with the right applications? Can it keep going until the job is complete? Can I review what it did?
That is why agentic systems matter. We are no longer just comparing models. We are comparing working environments.
Excel Plugins Are a Serious Productivity Advantage
One development I think deserves far more attention is AI inside Excel. This is no longer a toy.
There are now serious plugins from Microsoft, Claude, and ChatGPT. They are not equally good (Microsoft’s is the least impressive, Claude’s is probably the strongest), but they are all very capable. They can help with formulas, spreadsheet logic, data cleaning, restructuring, error-checking, and explaining models you have inherited from someone else.
If you are not using AI plugins in Excel, you are leaving a lot of time on the table.
Here is how I think about it. If a plugin reduces a day of work to an hour, you can then spend thirty minutes double-checking everything it did and still come out enormously ahead. That is not a hypothetical. I have seen people do exactly this. The time savings are real, and they are large.
No, it does not fully automate everything. Yes, you still need to verify the output. But that is not a meaningful objection. It is just good practice.
PowerPoint Is Much Less Impressive
PowerPoint is a different story. Anthropic now has a plugin for it, and it is better than Microsoft’s Copilot. But I still would not call it good. It has very limited capabilities, and if you are a serious PowerPoint user, I would not expect much from it.
I work in PowerPoint all the time. I use the plugin maybe once a week. That tells you most of what you need to know. Excel plugins feel like serious tools. PowerPoint plugins still feel like toys.
What This All Means
Claude CoWork is now the one AI tool I would keep if I could only keep one. Not because it is always the smartest, but because it gets closest to being a real co-worker.
ChatGPT 5.4 is very good. In some tasks it may be the best.
AI inside Excel is already a major practical advantage that too many people are ignoring.
And most importantly: we are moving from AI as assistant to AI as workflow.
That is the bigger shift. For a while, the conversation around AI focused on which model writes the best paragraph or scores highest on a benchmark. Those things still matter. But the more important question now is which system helps you actually work better.
The winners will not be the people with access to the single smartest model. They will be the people who learn to combine the right model, the right interface, the right files, and the right workflow. And that combination is not theoretical. It is something you can start building today.
This is not a future trend. This is how work is already changing. Exciting!
Created with
