For most of my time using Claude Code I’ve used it via Anthropic’s API rather than using a Claude subscription. I briefly tried the base Max plan but found myself out of usage almost immediately, and went back to API usage.
That usage has been steadily increasing as I’ve developed my workflows and tried different harnesses. It recently got to the point where I was regularly spending more than $100 per day on Claude tokens.
This is fine when most of that cost is passed on to clients. However recently one of our clients reduced the amount of work they’re sending us, meaning a higher percentage of that usage moved to non-billable internal projects. That proved to be unsustainable.
My brother has been using a Claude subscription without issue, so I decided to give it another shot. I’m now on the Max 20x plan, and it seems to be working well this time.
One unfortunate consequence of this switch is that I’m effectively no longer allowed to use third-party harnesses with Claude Code. So that means saying goodbye to Cursor 3 and Zencoder for the time being. Though Warp has recently added enough integration with Claude Code to at least give a hint of some of the things I liked about using third-party harnesses.
The biggest thing I’m missing so far is the ease with which Zencoder allowed me to create new tasks with their own worktrees. Claude Code does have its own worktree support, but it’s always felt more intimidating to me to get it to work well. I guess this is an opportunity to push through that and see just how good Claude Code’s worktree support can be.
It’s been interesting trying to come to grips with the incentives that this new way of doing software development sets up. As a more-or-less solo freelancer, I’m used to my productivity being almost entirely untethered from my wallet. If I want to produce more results, I can simply spend more time coding. No additional expenditures necessary.
Now that I’m using AI for programming, that’s much less the case. I can theoretically produce an unlimited amount of output, quality of my workflows allowing. But once I max out my subscriptions, that productivity additionally scales my costs. It’s more like having employees, where I could theoretically hire as many employees as I wanted to scale productivity, but it’s expensive to do so.
Except it’s not quite like employees.
- Employees are hard to scale up and down. You have to find them, interview them, hire them, and perhaps fire them. AI, on the other hand, scales up and down without friction.
- A new employee needs training and experience to realize their potential productivity at a task, whereas a cloned AI agent immediately has the same productivity as all previous agents.
- An employee is likely to be specialized in a few tasks, and moving them to a different task can be costly and disruptive. While AI resources are fungible, easily repurposed for whatever the current needs.
This creates a tighter coupling between a task’s profitability and the incentives to scale the execution of that task. With AI, if a task makes you money this month, you can spin up an unlimited amount of capacity to exploit that opportunity. And if next month it stops making you money, you can spin all that bandwidth back down to zero.
That’s different from scaling productivity with employees, where hiring many employees results in training costs, ongoing pay if you keep them on, and human suffering if you fire them.
I think historically businesses have simply split the difference. Hire enough employees to exploit most of the operation’s typical opportunities, and give them lower-value work during the times where those opportunities are less abundant. Accept some level of firing and hiring, but keep it manageable. Some lost opportunity, some wasted bandwidth, but hopefully not too much of either, most of the time.
AI promises to solve those issues. Perfect exploitation. Perfect efficiency. Societal costs still uncertain.