Expensive AI subscriptions are worth it now

When Open AI first launched its $200/month ChatGPT Pro subscription back at the end of 2024, I laughed. I didn't even sign up and test it for the article I wrote about it here on Zapier—it offered so little extra benefit over a $20/month ChatGPT Plus subscription that I didn't think there was much point. Wow, have things changed. 

Over the past six months, AI coding tools and AI agents have crossed an important threshold. They've gone from mostly doing what you want most of the time to almost always doing what you want almost all of the time. While the progress to reach this point was incremental, it represents a huge shift in what AI tools are good for, who can get value from them, and how you use them.

Eighteen months ago, I wrote on this site that ChatGPT Pro was basically for no one. Now I'm here to explain why I pay $200/month for Claude and consider it money well spent. 

Table of contents:

• The November shift

• Chatbots aren't the only interface anymore

• Coding tools turn anyone into a nerd

• Doing things better and doing more things

• Agents are adding value

• A few caveats

• AI is really having its moment

The November shift

As a writer covering the space, it's been clear to me for the last few years that AI tools could be useful when bolted on to existing apps. But until late last year, there hadn't yet been a killer standalone AI product capable of generating real revenue. 

ChatGPT was useful for some things, but OpenAI was (and still is) losing millions of dollars per month because not enough people are prepared to pay for a chatty version of Google. Similarly, AI image generation can be handy, but AI-generated movies and TV shows—or even AI photo shoots—aren't dominating the culture. OpenAI recently announced it was shutting down its video generation tool, Sora, and AI-generated imagery seems to mostly make headlines for all the wrong reasons. 

Simon Willison, one of the leading AI bloggers, calls it the November inflection point. Writing in The New York Times, developer Paul Ford pegs November 2025 as the time AI coding tools "suddenly got much better." It's when OpenAI released GPT-5.1 and Anthropic dropped Claude Opus 4.5—and when AI tools became really, really good. 

Of course, things had been building for a while. Claude Code was released in February, and its instant success among developers caused an industry-wide shift. In an interview on Lenny Rachitsky's podcast, Willison claimed that for most of 2025, OpenAI and Anthropic were almost exclusively training their models on computer code.

In November, though, Claude Code (and OpenAI's equivalent, Codex) went from being a tool you needed to be a developer to use to a tool anyone comfortable opening a terminal could get some value from. You don't need to have a deep understanding of every line of code to get somewhere; just be comfortable looking at it. 

And now, with tools like Cowork and computer use, and the friendly app-based interfaces of Claude Code and Codex, these tools are even easier for more people to use. 

Chatbots aren't the only interface anymore

AI chatbots were a great proof of concept. They packaged AI into a friendly and easy-to-use app, but there were inherent limits to the form factor. Because the AI was constrained to its chatty sandbox, there was very little it could do. Features like web browsing and research made them more useful as search competitors, but they were never able to do much. 

Coding agents are freed from many of these constraints. By being able to run terminal commands on your computer, they can operate on your files as well as chat back. In simple terms, you can use one to sort your photos into folders by date taken or rename a load of email archives. Of course, this kind of access means they can delete all your backups—but that tends not to happen if you're sensible. (I'll get into caveats a little later.)

More importantly, though, coding agents can write and run computer code—and they can now write it well. I'm one of millions of people who "learned" coding to the point where I could understand the general ideas without ever getting proficient enough to create functioning tools. I could mash together bits of PHP and JavaScript from StackOverflow to add features to my website and tweak open-source code to get it to run on my machine, but real tools were well beyond me. If you're like me, you can realistically now learn how to use Claude Code or Codex to build apps and automations you can use.

One incredibly important innovation here is Plan Mode. This is where the coding agent looks at what you want it to do and creates a plan for how it would do it—but, crucially, doesn't actually do anything. If you employ a bit of common sense, actually read what the AI is suggesting, and ask it (or Google) any bits you're unsure of, almost anyone can safely use a coding agent. 

Coding tools turn anyone into a nerd

I feel it's time for some concrete examples. Until recently, when people asked what I used AI for, I was able to say research, proofreading, and testing so I could write about the tools. I had a couple of good examples around translation, but for the most part, I couldn't make a strong case from personal experience that they were very useful. That's now changed. 

Over the last few months, here are some of the things I've used AI to do:

• Build a personal finance tracking dashboard and multi-currency invoicing tool that's a native Mac app. (No more QuickBooks for me!)

• Spin the invoicing part of my personal app into a real invoicing tool for freelancers.

• Build an SEO tool using DataForSEO that runs daily from my Mac.

• Completely rebuild my personal website. 

• Finally get my Home Assistant automations running properly.

• Build an automated series of filters for my Gmail account that keep the cold pitches and newsletter spam out of my inbox.

• Build an MCP server for my wireless thermometer so Claude can give me advice on how my barbecuing is going.

• Analyze months worth of chess games to find what mistakes I keep making. 

There have also been smaller projects, like migrating from one email service to another and cleaning thousands of orphaned and badly organized files on a media server, and larger projects that I'm keeping quiet for now that these tools have been instrumental in. 

And I'm not the only one. My partner's a lecturer, and she's been using ChatGPT to create Google Sheets automations that replace hours of manual copying and pasting. My editor here at Zapier has been using Cursor to automate some of her editorial workflows, fact checking and flagging issues that save hours of mindless Googling. 

Doing things better and doing more things

Looking at what I've used AI for over the past few months, I think there are two clear categories of tasks:

• AI has helped me do the things I was already doing better and more completely. 

• AI tools have allowed me to tackle projects I've always wanted to but didn't have the time or skills to do.

Take those Home Assistant automations and media server sort. I have enough technical chops to set up and manage a smart home without any more issues than normal. The problem is that going from "kind of working" to "actually feels like a smart home" requires dozens of automations with countless edge case rules built in. This is the kind of thing I'd have spent months tapping away at before, but Claude Code is able to take a prompt like…

Have the living room lights dim from 100% down to 60% between 9pm and 9:15pm except on Fridays and Saturdays when it should do it at 11pm to 11:15pm. If I turn the lights back on between then and 6am, they should stay dimmed. Otherwise, they should go to full brightness. 

…and turn it into a series of rules that work. If things aren't quite as I want, changing them just requires a simple plain language prompt, not a load of YAML.

And then there are the bigger projects like my website redesign and financial tools that were always on the to-do list, but I couldn't justify the time it took to figure them out. "Learn Swift" has been on my to-do list for around a decade, and I've never got further than installing Xcode. Here, coding tools are a massive accelerator. 

Agents are adding value

A year or two ago, the economics of all this didn't work. Developers could vibe code for fun, but regular people frequently got themselves into trouble, and building apps that actually worked was a challenge. 

Now, if you're motivated and are willing to learn some of the technical details, you can get a huge amount of value from coding tools—and that's before you even look at what's possible with agents that run on a schedule. 

In my three-month AI coding binge, I think I've saved myself money, despite being on Claude's $200/month Max plan. Building my own personal finance app that replaces QuickBooks will save me around $1,000 this year alone; throw in the articles I've written and the other efficiency gains, and I'm getting a very good return on the time I'm investing. (Plus, getting the smart home automations working properly has saved me a lot of interpersonal stress.)

And all of that is without any of the side-projects coming to fruition. 

A few caveats

The big caveat: use AI coding agents at your own risk. If you don't know what -rm -rf does, you shouldn't give them unfettered access to your system. Even then, the more access they have, the more possible it is for things to go wrong. I've not yet had any bad experiences, but plenty of people have.

This is where the Claude Code and Claude Cowork tabs in the Claude app as well as the Codex desktop app are so good. Not only do they give you a friendlier interface for working on your local machine, but you can write and run code on Anthropic and OpenAI's remote servers. It's a safer test-bed for whatever ideas you have. I first tested these tools by building plugins and automation flows for apps I was already using. Only when I trusted them a bit more did I give them any kind of access to my Mac. 

You also have to be very careful with important data. As a freelance writer, I don't really have to worry about sensitive documents and the associated legal ramifications of working with them. Common sense is key. If you want to use AI coding agents with business data, make sure your company has a service agreement. 

The cheaper Claude and ChatGPT plans also include Code and Codex; the limits are just lower. You don't have to go all-in to test it out. Pick something small, and try it on a cheap plan. If the AI agent idea clicks for you, then you can go further. But if you just want a chatty Google, the free ChatGPT plan is all you need. 

AI is really having its moment

A lot has changed in the last few months. Claude Code, Codex, Cursor, and other agentic coding tools have shown a real use case for AI. Not only do they enable skilled developers to work much faster, but they allow almost anyone to create working tools.

Of course, there's a huge danger factor to that. Some people will get in over their heads and accidentally expose financial secrets, personal data, and more. But included guardrails make it easy to safely test out these tools—and I'd recommend you do. I've been writing about AI for over a decade, and this is the first time I've really been able to say that it could dramatically change how you work.

And when you connect these tools to the rest of your software stack, they can take action across all the apps you use every day. Zapier MCP gives your AI tools structured, reliable ways to interact with thousands of apps—triggering workflows, moving data, and stitching together systems without you needing to manually wire up APIs. Or if you're a bit more technical, you can use Zapier SDK to build more custom solutions. 

This is the kind of thing that makes these expensive subscriptions start to feel less like a luxury and more like infrastructure.

Related reading:

• What is agentic AI? And how you can start using it

• Agentic AI vs. generative AI: Key differences and use cases

• Types of AI agents to orchestrate your workflows