I am spending too much money on AI.

Two subscriptions at home, three at work. Every couple of months I open the billing page with the intention of killing one, and every couple of months I close it without doing anything. This post is me trying to figure out why.

The setup

I’m a Google fanboy from way back, as I’ve mentioned more than once. Watch, phone, car, Chromebook, TV. That means Gemini is getting shoved at me at all times, whether or not I asked. Sometimes appreciated. Mostly interruptive.

At work we’ve been finding out about the limits of AI the hard way, which has produced our own little Skynet of self-replicating agents running business processes and saving hundreds of hours. (Those are their savings numbers. An argument could also be made that they invented enough new business processes to offset a chunk of that. We’re pruning.)

From my own desktop, separate from anything company-wide, I’m deep in ChatGPT, Claude, and Gemini. They handle administrative work, draft documents, do research, and comb through hours of meeting transcripts to pull out the three things that mattered. Wired up to other tools over MCP, plus the AI baked into most marketing platforms now, they also take on a good amount of client work that used to be pointing and clicking. (Aside: automating that stuff is quietly shifting client engagements toward strategy, where they used to get mired in production work for the first month or two.)

At home things are simpler. Until AI can do the dishes, my needs are modest. Though I do moonlight as marketing director for my wife’s business, which is where the requirements get real again.

Which model is actually better?

Nearly every AI comparison I run into is written from an engineering standpoint. SWE-bench scores, how cleanly it one-shots a React component, context window size, tokens per second. Which makes sense, because software developers are either the heaviest users of this stuff or at minimum the loudest, and the whole conversation is mostly them talking to each other. But it means the conversation measures things I almost never do, and skips right past the things normies do all day.

So this is the other version. Someone who writes, researches, drafts documents, automates the tedious parts of his own job, and does marketing on the side, sizing up three models on that basis.

What I’ve learned across work and home is that some models are just better than others at specific things. Long-form writing: Claude. Short-form writing: Claude. Business tasks involving calendars and documents: Claude. Making a decision inside a messy problem, like driving a browser to accomplish something: Claude.

You can see where this is going.

Gemini is embarrassingly mediocre at all of it, which is a strange thing to say about a product that was clearly ahead of the pack a couple of years ago. Ask it for a long document and it craps out around two pages no matter how hard you push on word count. Ask it for a blog post and it reads like someone mashed every business bro on LinkedIn into a style guide. Email isn’t much better.

And Claude Cowork, ahh, I love it. It’s really made for breaking down projects and work streams, running complex agents without me hovering over them, and reusing skills across jobs so I’m not re-explaining myself every Monday. An enormous amount of actual work-work gets done in there.

So here I am. Two subscriptions at home, three at work, Claude clearly the better model, and I still can’t bring myself to cancel Gemini. Why?

Because Gemini is playing a different game

Claude gets the nod on everything an LLM is supposed to be. Gemini isn’t really competing there.

Gemini makes pictures. Only occasionally useful, and the watermark makes it less so (you can work around that by going straight to the API), but highly entertaining. Yes, I acknowledge I am killing the environment to make silly pictures.

Gemini turns anything into a podcast. This sounds insane right up until you have a dense research topic and forty minutes of dishes ahead of you, and two peppy fake hosts give you the rundown. I’ve used it to prep for D&D adventures and to compress an interminable business-self-improvement e-book into an hour that contained the four good ideas in it.

Gemini builds apps on top of spreadsheets. The Canvas feature in Sheets. Sure, I could ask Claude to vibe-code something, connect it to Google Sheets, and get a similar result. But the output is pretty rough unless you understand data models and can write a decent programming prompt. With Gemini it was one prompt, one refinement, and I had a genuinely good scheduling optimizer.

Plus they’ve made it able to clean up a rambling voice note into a sendable text, build a home screen widget from a description, write three minutes of music, drop an interactive 3D model into the middle of a chat, and turn a research pile into an animated video.

The through-line is accessibility. Nothing to connect, nothing to host, no extra subscription.

Could I get Claude to do these things? Yeah, probably. I could vibe-code something that talks to a spreadsheet. I could wire up some text-to-podcast service over MCP. But I’d have to do it. I’d have to connect it. I’d have to buy subscriptions to all of it. And the vibe-coded spreadsheet app is an hour of bug fixes and data model refinement before it works, plus somewhere to host the thing when it does.

So here’s where I’ve landed

Gemini is there to do stuff. It watches my calendar and summarizes my inbox. It makes a picture of my cat with butterfly wings to send to my nephew. It searches across everything Google knows and turns the result into a podcast or an infographic. It tells me how well I slept last night (note to self: turn this off, I can generally assess my own restedness by standing up). In Android world it apparently also makes watch faces and widgets and drives my phone.

Google has cleverly gated the good stuff behind a subscription, but that subscription comes with generous usage limits, a pile of Drive storage, and better features across a bunch of Google apps. It’s not a bad deal.

And Gemini will also do LLM stuff. It’ll write a short document, a generic email, a LinkedIn post. It does okay. Not great, but fine. The job gets done with all the skill of a high school intern.

It’s also fast as heck. Even the Pro model makes Claude feel creaky. It’s very possible that latency is a big part of why Claude reads as smarter.

If I weren’t doing work for my wife’s business, where I need something with enough taste to update a blog and enough precision to write SQL against a data warehouse, I could get by on Gemini alone.

But I do need that extra sharpness here and there. So two subscriptions it is, at least for now. I may start cancelling Claude and re-upping it in the months I actually need it, which feels like a petty little victory I’ll enjoy for about one billing cycle. And I’ll keep an eye on Spark to see whether it ever gets close enough to be my at-home Cowork substitute.

And ChatGPT?

Zero need. I’ve run head-to-heads, and it beats Gemini at some things and Claude at others. It isn’t better in any way at all than the combination of the two, and I’m already paying for both.

That subscription, at least, I can cancel.