If you want AI to do real work and not just answer trivia, pay $20 a month for Claude or ChatGPT and use its agent mode on a task that matters to you. That’s my advice for this fall. The rest of this piece covers how to choose between the two, what changed this year, and what to watch out for.
The question I get most often about AI is also the simplest one, which is which tool to use. People ask it at lunch, after talks, and in the hallway at work. Most of them have tried a free chatbot once or twice (I’m looking at you Copilot) and weren’t impressed, knowing they’re missing something.
Fact is, they usually are. The tools changed a lot this year, and the free versions most people tried are not the ones doing the interesting work. This is my current answer as of October 2026. I’ll update it when it stops being true, which at the current pace may not take long.
I’ve spent my career in the space between networks and the people who rely on them, first as a sales engineer and consultant working out what customers really needed, then leading growth at a rural fiber provider. Somewhere along the way I became the person who builds the tool instead of waiting for one, and these days that mostly means figuring out where AI actually helps a small team.
The short answer
Pick Claude or ChatGPT, pay the $20 a month, and use it on real work for a few weeks.
That is close to what Ethan Mollick, who teaches at Wharton and writes the best regular guide on this, concluded in his summer 2026 edition. I agree with him. For casual, low-stakes questions like a recipe or a birthday card, any free chatbot is now good enough. The gap shows up when the work matters: a contract you want a second read on, a spreadsheet you need analyzed, or a project that takes more than one back-and-forth. There, the paid tiers run the strongest models, and the difference is large.
Why those two and not the others? They are the only two that combine a top model with a good way to let it actually do work for you. Google’s Gemini led not long ago and could again. Today, though, its newest model is limited to a small group of government and security users, and its everyday tools trail on the kind of multi-step work described below. Microsoft Copilot is fine for drafting inside Word and Outlook if that’s what your employer provides, but it lags behind on the rest.
Choosing between Claude and ChatGPT
For most work they are close enough that the best one is the one you will actually use. The differences that matter are in what surrounds the model.
| If most of your work is… | Start with | Why |
|---|---|---|
| Writing, editing, long documents | Claude | Anthropic’s September release focused on clearer writing that leads with the point and follows your style rules |
| Images or graphics | ChatGPT | ChatGPT has a strong image generator built in. Claude has none |
| Talking it through out loud | ChatGPT | Its voice mode listens and speaks natively, so it handles interruptions like a real conversation |
| Research across a pile of sources | Either, plus Google’s Gemini Notebook | Mollick calls Notebook (formerly NotebookLM) the most useful research interface for analysts and writers |
| Word, Excel and Outlook at a Microsoft shop | Copilot, if that’s what you’re given | Good for office documents, weaker at multi-step work |
| Building software or automating your computer | Either: Claude Code or OpenAI’s Codex | Both are strong. Try each on the same task |
If you can afford it, run both for a month. Give them the same real task and see which result you would rather hand to a colleague. If you’re doing image generation, start to think about specialized tools like Canva or Adobe Express. Your company may already have a subscription you can use.
From chatting to delegating
The biggest change this year is not a smarter model. It’s that AI tools now do work instead of only answering questions. The industry calls these “agents.” In practice it means you hand the AI a task and a set of tools, and it goes off and works for several minutes, or an hour, before coming back with a result.
Both companies now offer two ways to do this:
- On their computers. Claude calls this Cowork and ChatGPT calls it Work. You describe the task, connect the apps you’re comfortable sharing (email, a calendar, a folder of documents), and the AI works on a computer the company provides. You can start a job from your phone and check back later.
- On your computer. Claude Code and Codex run on your own machine. They can work through many files over a long session and can operate your apps the way you would. This is more powerful and needs more care. Recent updates even have options of ‘cloud computers’ where they can run without your PC – that’s an advanced topic for another time.
Mollick gives a good example. He asked both tools to read his email and prepare for a seminar he was teaching. About 10 minutes later each had built teaching materials and drafted a reply to a colleague. That would have taken him a couple of hours.
The skill this asks for is closer to managing than to prompting. Say what “done” looks like, look carefully at what comes back, and ask for changes the way you would with a new employee. Anthropic’s own advice for its latest model is roughly the same: hand over the whole task, define the finish line, and agree on when it should check in.
This is the point where you’re likely excited about digging in but take a breath and focus your goal. Write out what you’re trying to accomplish, why it’s important and what the “good result” looks like in the end. Using Claude or ChatGPT as a tool means giving it boundaries, rules, structure and goals.
What it costs
For an individual, the price hasn’t changed: about $20 a month for either service, with higher tiers for heavy use. What has changed is how much work that $20 buys.
On September 22, Anthropic and OpenAI released new models within an hour of each other. Both put lower prices front and center, according to that day’s AINews recap:
| Model | Price per million tokens, input / output | Change |
|---|---|---|
| Claude Opus 5.5 | $4 / $20 | About 40% cheaper per task than the previous Opus at default settings |
| GPT-6 Sol | $2 / $10 | About 50% cheaper than the model it replaced |
| GPT-6 Luna | $0.10 / $0.50 | About 50% cheaper, built for high-volume simple work |
The research group Epoch estimates the cost of AI at a fixed level of performance has been falling about 47% per quarter since 2023. There is a catch: newer models often “think” longer and use more tokens (a token is roughly three-quarters of a word), so the sticker discount doesn’t always reach your bill. For one benchmark run at maximum effort, independent testers found the savings disappeared entirely.
If you’re a business owner, the takeaway is that cost has stopped being a good reason to wait. If you decided a year ago that some AI project was too expensive, run the numbers again.
Three cautions
Keep the approval switch on. Both tools let you choose whether the AI asks before it sends, buys, deletes or changes something. Mollick found this out the hard way. In his seminar test, Claude drafted the reply to his colleague, but ChatGPT sent it, because he had earlier given ChatGPT permission to send email. Leave approvals on until you understand how the tool makes mistakes.
Be careful what you connect. An agent that reads your email or browses the web will run into text written by strangers. Some of it is written to trick AI assistants into doing things, such as forwarding files. This is called prompt injection. The models resist it better than they used to, but the problem isn’t solved. Connect only what the task needs and only from reputable sources.
Put a ceiling on spending. You might be swayed by a friend or work buddy to move up a level to a ‘max plan’ or similar. Don’t. Since you’re starting out, you need to learn how to make the best use of the resources you have because it’s a required skill later. Once you move past the $20 plan into anything billed by use (aka API billing), set a hard limit. Simon Willison, one of the most careful independent testers of these tools, put it plainly after OpenAI’s developer conference: we will need default hard budget caps on pretty much everything. An agent that runs for an hour can run up a bill for an hour too.
And for anything with real consequences, such as medical, legal or financial questions, use the strongest model you have access to and treat its answer as a second opinion, not the last word.
Where to start
Pick one task you already do every week that takes an hour or more, such as a report, a proposal, or cleaning up a messy spreadsheet. Pay for a month of Claude or ChatGPT and hand it that task in its agent mode. Look closely at what comes back and ask for changes until it’s right. You’ll learn more about what AI means for your work from that one experiment than from any guide, including this one.
That’s it for now. If a colleague is wrestling with the same question, feel free to forward this along.
Sources
- Ethan Mollick, An opinionated guide to which AI to use to do stuff, One Useful Thing, July 2026
- Latent Space, Claude Opus 5.5, the new default model for AINews, September 23, 2026
- Simon Willison, OpenAI DevDay 2026 live blog, Simon Willison’s Newsletter, October 5, 2026
Leave a Reply