The Extension That Wasn't a Gift

Spent an unreasonable chunk of my Sunday evening reading a comment thread about Anthropic extending Claude usage limits for paid users until the 12th of July. Not because it affects me much, I’m on a modest plan, but because the whole thread was a perfect little snapshot of where we’ve landed with AI tools. People thanking a trillion-dollar company for temporarily not taking something away.

The extension itself is simple enough. Fable 5 (that’s what everyone’s calling the model, apparently) was due to hit its usage caps, and Anthropic pushed the deadline out a bit. Cue relief, cue suspicion, cue at least one person accusing another of being a corporate shill for pointing out that compute costs money. All in the same thread. That tension never really resolved, and I don’t think it’s supposed to.

What got me was a comment further down: someone describing how their “work slowed down dramatically” the moment they hit their limit, then sped right back up after they paid fifty bucks for more tokens. Reads like a caffeine withdrawal story, except it’s a work tool, and the person saying it clearly wasn’t joking. I’ve felt version of that myself, minus the fifty bucks, when a CI pipeline goes down and I realise how much of my thinking now happens through a prompt box rather than in my own head. Not proud of it. Not deleting it either.

There’s a genuine debate buried in the noise, and it’s one I don’t have a tidy answer for. One side says compute isn’t free, subscriptions have always been subsidised, and companies are allowed to adjust pricing as costs catch up with reality. Fair enough, that’s just business. The other side points out these are trillion-dollar valuations we’re talking about, built partly on training data none of us consented to, and consumers aren’t obligated to care about a corporation’s margins. Also fair. I lean toward the second view most days, but I’m not going to pretend the first one is nonsense.

What actually unsettled me was the bit about people running ten separate accounts at $200 US each just to keep using one model without limits. Someone in the thread called it “crazy shit,” and, yeah. I work in DevOps, I’ve automated things I probably shouldn’t have, and even I found that a bit much. Somewhere between “useful tool” and “load-bearing infrastructure for your entire workflow” there’s a line, and a decent chunk of that thread had crossed it without appearing to notice.

I think about this stuff more than I used to, partly because of the job, partly because my daughter’s generation is going to inherit whatever version of this we settle on. I’m fascinated by what these models can do; genuinely, unreservedly fascinated, the kind of excitement I haven’t felt about a piece of technology since I first got broadband at home. But I also can’t shake the environmental cost of all this, the water and power going into training runs and inference at scale, while we’re bickering in comment sections about whether our free samples got smaller.

None of this is going to resolve itself neatly, and I don’t think it should. The compute is real. The margins are real. The frustration of hitting a wall mid-task is real too, and so is the discomfort of watching people treat a chatbot subscription like oxygen. All of it can be true at once.

For what it’s worth, I’ll keep using these tools, carefully, and I’ll keep being a bit uneasy about how normal that’s become. Feels like the honest position, even if it’s not a satisfying one.