I use Anthropic’s Claude AI tools for very different jobs: How to pick between models, Code, and Cowork
Follow ZDNET: Add us as a preferred source on Google.
ZDNET’s key takeaways
- Claude’s lineup spans chatbots, agents, and AI models.
- Code builds software. Cowork handles broader computer tasks.
- Agents can be force multipliers but require close supervision.
Anthropic is currently the world’s most valuable AI company, with an astonishing $1.2 trillion valuation on secondary markets as of earlier this month. The company’s offerings have been making headline news all year, especially after the US government temporarily banned use of Anthropic’s Fable model because it was just too smart and risky for our own good.
But cutting past the hype, the excitement, the hyperbole, and the financial insanity, what Anthropic products might you and I care about? What are Claude.ai, Claude Code, and Claude Cowork? For that matter, what are Haiku, Sonnet, Opus, Fable, and Mythos?
Also: Claude Fable 5 is back, but I’m sticking with Opus 4.8 for daily work: 5 reasons why
In this article, we’re going to dig in. We’re specifically going to look at Claude.ai, Claude Code, and Claude Cowork, but we’ll also put Haiku, Sonnet, Opus, Fable, and Mythos into perspective.
The motor vehicle analogy
Sometimes, I find it easier to understand something if I can relate it to cars. So let’s do that with Anthropic’s offerings.
When you walk into a car dealer, you typically make two sets of choices. You choose the vehicle model. Do you want a regular sedan? Do you want an SUV? Do you want a pickup truck? They all have their pros and cons, and they all provide advantages for certain types of activities.
Also: Claude Sonnet 4.6 delivers frontier-level AI for free and cheap-seat users
The second choice is how you equip that SUV or sedan. The key choice is how big an engine you put in the vehicle. I have a Dodge Challenger RT, a very nice muscle car. The Challenger model ranges from something that makes for a very peppy personal daily driver all the way up to a competitive race car. The difference is mostly in the engine and suspension.
My car, for example, came with a 375-horsepower, 5.7-liter Hemi V8. That’s powerful — powerful enough to have a race mode that I’ve never been brave enough to try. My car even has a racing stripe, which I like to think gives it better aerodynamics (but probably doesn’t).
But my car’s engine is nowhere near the 6.2-liter supercharged High-Output Hemi V8, which turns the Challenger Demon into a street-legal monster that can handle extreme drag-strip duress. With three times the horsepower and double the torque of my car’s Hemi, the engine is a massive performance improvement over what I drive.
But here’s the thing. I like my car to have some pick-up-and-go. I like it to growl a little bit. But I’m not taking it onto the drag strip or the racetrack. I’m just not. I drive to the hardware store and the coffee drive-through. I work from home, so I don’t commute with it. I can’t technically even call it my daily driver because there are days I don’t even fire it up.
That brings us back to AI. Haiku, Sonnet, Opus, Fable, and Mythos are all LLMs (large language models), roughly equivalent to the different engine sizes. When you use the car analogy, you begin to see why Fable and Mythos were banned. They’re like the racecar engines. They are absurdly powerful and much more expensive. Opus is like my nice 375 hp V8. It’s powerful, but not too powerful. And Haiku and Sonnet are like the smaller V6 engines that are available for some of the Dodges. Good, but not great.
Also: I compared Claude Opus 4.8 with 4.7 in a 10-round honesty test – and a legal prompt broke it
Next, think of Claude.ai, Claude Code, and Claude Cowork as different vehicle types, used for different things. Claude.ai is the chatbot, while Claude Code and Claude Cowork are the agentic helpers that carry out tasks on your behalf.
The agentic Claude Code and Cowork are more like pickup trucks or panel vans than basic cars. They get work done. They can even be specially outfitted with skills and plugins, roughly analogous to how your work van might have a ladder mounted on top and be fitted out with tools and specialty gear inside.
Chatbots vs. agents
Put simply, chatbots answer questions while agents do stuff. Now, clearly, that’s quite an oversimplification, but it’s a good starting point for our discussion.
Back in 2023, when generative AI first hit the scene, all we were doing was feeding prompts and questions into chatbots. The AIs answered questions by searching their considerable knowledge bases. But there were limitations, because their training data had cutoff dates, after which the AIs had no information.
Also: The best AI chatbots of 2026: Expert tested and reviewed
Then, a year or so later, the chatbots started to be able to incorporate live web searches into their answers. That, technically, changed how the AIs “thought” because they were using two information sources: their training data and the whole web (or, at least, the parts of the web where they could gain access).
They started to be able to carry on conversations, keep track of information from previous conversations, know something about you that they retained between sessions, and perform work like programming, but within the limits of giving you chunks of code you could cut and paste into your programs.
ChatGPT, Claude.ai, and Gemini are all chatbots. The branding gets a little mixed because Gemini is used by Google in its marketing to refer to its AI regardless of where it’s running.
By contrast, both Anthropic and OpenAI use their chatbot names followed by focus words to specify their agentic tools. So ChatGPT is the chatbot, while ChatGPT Work and ChatGPT Codex are the agents. Likewise, Claude is the chatbot, but Claude Code and Claude Cowork are the agentic tools.
Also: I let ChatGPT Work and Claude Cowork loose on my files – only one made me nervous
Let me give you another analogy. Back in the days when we actually used phones for talking, I had some friends I’d talk to on the phone and have very pleasant conversations with. But I had other friends who would come over at 3 a.m. and help rewire my apartment building to provide power to my server rack.
Think of my conversation friends as being in the chatbot category. Think of the friends who would help you rewire an electrical box or help you move apartments as agents. Chatbots talk. Agents can carry out complex tasks across many steps.
That brings us to Claude Code and Claude Cowork.
Claude Code
Claude Code is probably the best known of the coding agents, although most of the big providers now have their own tools that can code. I’ve discussed Codex from OpenAI often here in the pages of ZDNET. Both do their coding jobs extremely well.
Also: Claude Code made an astonishing $1B in 6 months – and my own AI-coded iPhone app shows why
The idea of agentic AI is that agents can plan and execute a sequence of tasks, make decisions mid-project, course-correct, and then continue to the completion of a goal. In the case of Claude Code and Codex, rather than writing one block of code and printing it to the chatbot’s output screen, they create and modify directories, create and modify files, run the files, check results, fire off shell scripts, generate images, and even write and run their own programs.
They are both astonishing and far from perfect. They need strong guidance. I have given Claude Code assignments that would have required a year of personal development to accomplish in terms of building complex image analysis and transformation capabilities. But it still makes boneheaded moves. I have to tell it both how to use those capabilities and what capabilities to develop.
Also: I used Claude Code to vibe code a Mac app in 8 hours, but it was more work than magic
That said, with my constant involvement, they have built big, complex programming projects. More to the point, I have gone from ideation (just having an idea that I wanted something) through to completion in a fraction of the time the same work would have taken had I coded it myself.
Keep in mind, though, that I have experience here. Not only am I a former computer science professor, but I’ve also been a product marketing director and engineering manager. That means I know how to manage products and engineers as well as I know how to program.
I ran a software company for 15 years. Eventually, I transitioned from managing people and making payroll to a career as an author, teacher, and advisor. But those management skills are directly transferable to agentic management.
Also: I used Claude Code to vibe code an Apple Watch app in just 12 hours – instead of 2 months
Working with Claude Code is very much like talking to a team member over Slack. I’ll give it an assignment or an assessment and set it loose to work on a series of tasks. In fact, as I’m typing this article on my keyboard, Claude Code is currently working its way through an assignment to code a challenging image analysis and transformation capability.
As a solo developer, I find that being able to pair program with Claude Code or Codex is quite powerful and fundamentally freeing.
Claude Cowork
Claude Code and Codex are coding agents. Their primary workflow, interface, and feature set are designed for creating coding projects.
Both Anthropic and OpenAI have also opened up productivity workflow agents whose feature sets are more attuned to computer use, whether that’s organizing files or creating presentations. For Anthropic, that’s Claude Cowork. For OpenAI, that’s ChatGPT Work.
Also: I gave Claude Cowork 7 non-coding jobs, and it earned a spot in my toolbox
My classic example is using Cowork and Work to organize my PDFs folder by subject. While traditional utilities can easily sort files by date and file type, it takes an AI to determine categories and create a taxonomy. I’ve used both Cowork and Work to rename files to be more descriptive, determine a set of folder categories, and then move my PDFs into those folders according to each file’s content.
Workflow agents like Cowork and Work can also (with your permission) control your computer, use your browser, execute programs, and generate documents.
Both of these tools grew out of their coding counterparts. For the coding agents to be fully capable, they had to be able to create and move folders, create files, execute programs, and even control the computer.
After a while, managers at Anthropic and OpenAI noticed that some users started using the coding agents to do decidedly non-coding tasks. It was only in 2026 that both companies put a more user-friendly working environment on the workflow agents, making them more approachable for people carrying out non-programming tasks.
In another example, I had Claude Cowork dig through a week’s worth of email (that week, there were more than 7,000 messages) and find those that had specific opinions on the topic of an article I was writing. That sort of task would usually take me half a day or more, during which I would become more and more cranky. By delegating it to the AI, I got what I needed. I was substantially less insane than I would otherwise have been by the end of the day.
Recently, I’ve found that a particularly powerful combination is using one or more coding agents along with a workflow agent to do research and analysis. I really do feel like I’m working with a team.
The elephants in the room
With that, I’d like to discuss the two big elephants in the room: job replacement and security.
There is no doubt that AI agents are impacting jobs. They’re also lowering the quality of output. That’s why we see so much AI slop online, on YouTube, and in books on Amazon.
But for solo practitioners like me, they’re force multipliers. When I use an AI to augment my work, it’s not that I would otherwise hire a programmer or an assistant. I would either just do the work myself, taking much longer and stressing my wrists more, or I would simply not undertake the task.
Also: The top AI fear for 6,000 tech pros isn’t losing their jobs – it’s more work for the same pay
For nearly all the recent programming tasks I’ve been working on, it simply would not have made sense or have been justifiable for me to do that work. But with Codex and Claude Code, I launch them and guide them in and around the rest of my work. The results are some amazingly powerful and deep pieces of software, both intended for possible sale and just to solve vertical challenges I have in my own productivity.
Plus, if you ask the AI to build something, change your mind and have it remove what you just asked it to build, and then later that day ask it to build it again, it doesn’t glare daggers at you or walk around the office for a week pointing and muttering under its breath. Not that such a thing has ever happened to me. I’m just sayin’.
Also: I tried Claude Cowork on my Gmail inbox after Gemini choked – and it saved me hours of work
It’s the same with using a workflow agent as an assistant. I used to have AAs (administrative assistants) when I worked for bigger companies. But now, it’s me. That email sifting project still would have happened, but it would have been a very tedious half-day that, while necessary for the article I was writing, would have also been a poor use of my time.
Next, let’s talk about security. I am very uncomfortable letting the workflow agents loose in my email, my browser, and on my desktop. Right now, I only open up computer use, browser use, or email access on an assignment-by-assignment basis. Once the assignment is done, I remove that access.
Also: 1Password’s new Agentic Mode lets Claude log into your accounts without seeing your credentials
I still worry that the AIs will nuke something, so I’m also pretty scrupulous about backups. This is why I’m very hesitant, even with 1Password’s secure access assist, to open up most of my cloud applications to the AIs. It’s much harder to restore cloud application data than it is to restore backed-up desktop application data. So if one of the AIs ran amok inside something like QuickBooks Online, undoing the damage could be a herculean task.
And that’s before looking at whether to let the AI companies have access to my data, which is its own bucket of worry. Right now, my approach is to do some trade-off calculus. Will I save enough time for it to be worth the risk? Is the task I have to do so annoying that I’d rather take on the risk than the task? Or is the task interesting enough that even though there are some privacy and security concerns, I might get a cool article out of it?
For those situations where the answer is yes, I’m hesitantly beginning to experiment with the workflow assistants.
AI economics
There’s one final issue to keep in mind. Agentic AI uses a lot more AI juice than chatbot AI. That directly relates to cost and capacity.
All the AI companies limit usage based on purchasing tier. Free users get chatbot use, but with lower-powered LLMs and with less thinking time, as measured in tokens.
Users who buy the typical $20-a-month mainstream plan can get about as much chatting as they want but find themselves throttled when it comes to agentic use. I have been repeatedly throttled after about five hours of agent use on my $20-a-month plans.
This applies as much to workflow agents as coding agents, especially if you want the workflow agents to be active and monitoring stuff all the time. You can schedule agents to regularly take action, but doing so will use up capacity, possibly forcing you to purchase a higher tier of use.
When I was doing a big development push using Codex, I paid for OpenAI’s $200-a-month plan. That was hard to absorb. Now, I pay for the ChatGPT Plus plan at $20 a month.
I split up my coding work between agents based on project. Right now, I’m working on a bunch of projects that use Claude Code, so I’m paying for its $100-a-month Max plan. I use it a lot, and have found I rarely get throttled. But I am concerned that if I also use Claude Cowork on the same plan, I’ll be doubling up on my AI resource usage. That could impact availability.
What will you use?
This is all so much of a moving target. Home lab users are incorporating some of the less powerful but free models into their own home-based AI computers. Competitors are fighting to offer more and more for less and less, while other companies are trying to raise how much you spend each month by making their systems faster (for a fee) or smarter (for a fee).
Vendors are also moving toward unified, dedicated apps. The Claude desktop app lets you switch between a chat/Cowork mode and a Claude Code mode. I use the desktop app fairly regularly for Cowork, but I find it too limited for coding.
Also: How I set up Claude Code in iTerm2 to launch all my AI coding projects in one click
For some reason, while you can have multiple sessions in Code mode, you can’t have multiple tabs. I use multiple tabs and color-coded sessions in iTerm2. Until the desktop app can reproduce those simple features, I’ll continue to program using the terminal.
That said, AI-based coding has become such a huge part of my life that it’s hard to believe I’ve been using it for less than a year. I’ve derived at least six or seven years of personal programming productivity from it.
We’ll keep on looking at these tools, testing them out, and reporting back to you about them. But I’d like you to tell me if you’re using them and how you’re using them. Let us know in the comments below.
You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter, and follow me on Twitter/X at @DavidGewirtz, on Facebook at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, on Bluesky at @DavidGewirtz.com, and on YouTube at YouTube.com/DavidGewirtzTV.
