Reliance on an engine that runs in the cloud
This week Anthropic took a lot of its users by surprise when a temporary 50% increase to Claude Code usage limits came to an end. The increase had been running for months, but plenty of users seemed unaware it was temporary, or even that the extra capacity was there in the first place.
As someone who spends a fair bit of time doing front-end work, Claude is genuinely impressive. It has an ability to visualise concepts quickly and execute on an aesthetic that I find other models can struggle to see. Anthropic also ultimately gets to set the price for its product. With investors now talking about a potential $2 trillion valuation, there is a hell of a lot riding on what comes next.
GitHub is another recent example. In June it moved Copilot from largely flat-rate plans to usage-based billing, with monthly included usage before additional consumption is charged. Some users subsequently reported dramatic increases in what the same kinds of workflows could cost once those included credits were exhausted.
It feels like the period of throwing increasingly generous access at users to drive adoption is beginning to change. Cost is only one side of relying on these products though. We also can’t necessarily rely on the thing we’re paying for behaving the same way tomorrow.
In April 2025, OpenAI rolled out an update to GPT-4o intended to improve its personality and usefulness. Instead, the model became noticeably more agreeable and flattering, sometimes validating users when it shouldn’t. OpenAI ultimately rolled the update back. Same GPT-4o product name; noticeably different behaviour underneath it.
Earlier this year, Anthropic ran into something similar with Claude Code. A change to its default reasoning effort reduced it from high to medium in an attempt to improve latency. Users reported that Claude suddenly felt less capable. Anthropic also identified a context-handling bug that made it seem forgetful and repetitive, and a system-prompt change intended to reduce verbosity that hurt coding quality. All three were eventually fixed or reverted.
With variables like these, it isn’t hard to see why AI adoption inside businesses needs to become more nuanced.
Early on, it was common to hear AI described as a “black box”; mysterious and unknowable. That stigma has largely faded as AI has become part of everyday technology discourse. But there is another kind of black box worth thinking about when these tools become part of real pipelines and processes.
Even looking purely at dollars and cents, how heavily should a business rely on an engine it doesn’t control when it can’t be completely sure how many cylinders will be firing tomorrow, or how much fuel it will get for its dollar at the pump?