A while back there was a post on Hacker News about whether a $48K GPU server was worth it. Buried in it was a line I keep coming back to:
The mentality shift of renting vs. owning the gpus is huge. When renting, each experiment costs money and I had to ask myself is it worth it. When owning, it feels like not running experiments is costing me money.
My comment on HN pointed directly to this line:
I feel like there is some very deep generalizable wisdom buried here.
The same thing happens one level up, to the rest of us who will never buy a GPU. It's between subscription (rent) and API (buy) prices.
For the most part, I've been using AI on a subscription. Today, I often move between Anthropic Claude and Open AI Codex. In that mode, every idle hour feels like waste. I've paid for the subscription and I want to max it out. At the end of the day I find myself kicking off larger refactoring tasks, just to see if it can be landed.
However, you can also work in an environment with an Enterprise (aka API) license. There, each prompt racks up a cost. In this mode, you think about every prompt as you can see it ticking away. You're on the meter. You start to do cost-benefit for each prompt.
This is quite an adjustment as I'm used to throwing everything I can into a potential PR. Straight to Claude and Battleship Prompts are a couple of examples of this behaviour.
Neither of these is a rational response to price. They're both a response to framing. It may well be that metered pricing is a lot more cost effective. Or you might be using your subscription to generate unnecessary work for yourself.
Metered pricing is sometimes a choice, but for many enterprises it effectively becomes mandatory to get the security and controls organizations need[1]. There is the argument (probably a strong one) that the overall pricing nets out over different usage levels. In itself it's the rational position.
What I would suggest is that being on the meter can work against you. If someone is looking to learn, they might benefit from a subscription. Consider allowing employees to expense a private subscription - not for the company work, simply for open source or their own interest.
It means they can kick the tires, try different prompts, and learn the tools "off the clock." It could also let them try out different tools: you may have standardized on an AI-provider, but the offerings are changing very quickly.
Pricing drives behavior and that behavior may be good, bad, or indifferent depending on the context. As AI costs rack up, organizations are becoming much more aware of the ROI. It'll be back on the vendors to come up with packaging that will land all of this[2].