ChatGPT or Claude?

Last week, at an investor conference in Southern Europe, I ran an AI roundtable with a room full of people who had flown in from around the world to spend a few days in the same place. The discussion was one of the better ones I've been part of this year, and over the course of it I fielded dozens of questions about agents, about jobs, about where the money is going. The question though that stumped me was the most innocent one:

"Should I use ChatGPT or Claude?"

It should have been the easiest question of the day for me, and I paused. I have several accounts with each of them, partly because I keep hitting usage limits and partly because I stopped choosing a long time ago. I use them interchangeably. Most days I have them using each other, one agent reviewing what the other built overnight, a second model brought in when the first one confidently announces it has fixed something that is still broken. I'm so far inside both that choosing between them felt like being asked to name my favourite child.

So I said the honest thing. "With all due respect to everybody here, it doesn't matter which of these tools you ask. Both of them have far more intelligence than any question you're going to bring them."

I don't think I would have said that six months ago though.

There's a line I believe we've crossed with this generation of models, Opus 5.5 from Anthropic and Astra 6 from OpenAI. I'm burning one billion tokens a day now, across every tool I can get my hands on, and for most of the last year my complaints about the models were real ones. Context windows running out halfway through a session and leaving me to re-explain an entire project to an agent with no memory of the last four hours. Models that were good but not great, capable of getting me most of the way and then stalling. When a project died, I could usually point at the model and be at least partly right. Lately, when I'm honest about where my projects stall, the model is rarely the reason.

For most of my career, picking the right platform was one of the most consequential decisions a company could make. I started a mobile apps company twenty years ago in BlackBerry's backyard, at a time when which phone we built for could decide whether we still existed three years later. The whole culture around these tools is built to make the choice of platforms feel like the decision that matters. It's doesn't.

Earlier this year, I wrote about how every excuse I'd relied on for twenty-five years as an entrepreneur had evaporated with vibe coding. Before, I would point the finger at the engineers, our clients or even investors whenever I was feeling frustrated.

Years ago, my excuse was my team.

Months ago, my excuse was the model.

Now, the only thing left between me and my ideas, is me.

That realization has pulled me back into something I thought I'd left behind. I'd spent the last few months believing I had come out the other side of my AI psychosis, the late nights and the inability to stop. I can feel myself being dragged back in now. The difference this time is the energy underneath it. Knowing that I'm the limitation is humbling, and it's also producing a drive and a creativity I haven't felt in decades. My wife has seen a version of this before, and I'm not sure yet whether this one is healthier or the same thing with better reasons. I'd like to think I can tell the difference. I'm not certain that I can.

The tipping point happened quietly enough that most people missed it.

For years, every model release mattered to us in a very practical way. A better model meant a project that had stalled might start moving again, or a bug that had survived a dozen agent runs might finally give up, so I followed every launch closely. Somewhere in the last few months that stopped being true, at least for most of what most people do.

When one agent hands its work to the other for review, the feedback comes back in a different tone and style, and almost never catches something the first one couldn't have caught. If I can't reliably tell which one did the work, I'm sure the choice between the two doesn't matter.

For most people, there is now more intelligence available than they will ever ask for. The investor drafting a memo, the founder sketching a product on a flight home, the marketer running ad campaigns, the trader analysing stock trades. The frontier still matters to researchers and to the people building at the very edge, and I'll keep watching it, because I'm one of the people leaning in. But for nearly everyone else, the gap between the two best models is now smaller than the gap between the train and its platform.

As intelligence gets commoditized, the scarcity moves to whoever is typing: me.

What do I know that's worth handing to something this capable? Can I describe what I want clearly enough that it has something real to work with? The question is not about how to build, but what to build. And the tool question keeps me from finding out how far my own ideas can go.

Either model could take me further than I've been willing to go so far. What am I still waiting for?