Recently GitHub announced a change in pricing model for CoPilot which they’ve tried to wave away as not a price increase (because the list prices don’t move) but which really is as they are now charging for direct usage rather than per request.
What’s interesting to me isn’t so much the changes themselves - CoPilot had an exploitable model and most users of AI are doing so under heavy subsidy - but the questions people have started to ask as a result.
At the sub-$100 a month standard fixed cost model it’s fuelled adoption without any real barriers and as a result nobody has really applied any rigour to it. AI got something wrong? Run it again. Is the money we’re spending really increasing productivity? Well it’s a rounding error in a budget so no one looks.
But these changes shift the model in two key directions:
- Individual developers could be generating bills of hundreds, or even thousands, of dollars each. You can get some insight into how usage based billing lands by using Claude Code with AWS Bedrock - the bill you get, in comparison to a Max subscription, can be 5x plus based on my experience.
- The pricing is effectively no longer fixed. It varies per developer and with usage and it’s nigh on impossible to predict how many tokens an AI will use to fulfill a request: heck ask the same AI the same question and you can end up with really quite different token usage.
Suddenly this isn’t a business impulse purchase. It’s a line item — and line items get questioned. How do we control these costs? Are we getting value for money? How are you measuring whether you are getting value for money? What happens when one of these agents spends $500 and generates garbage? Who is liable when it goes wrong?
In fact if you look at the GitHub discussion on this you can see these questions already starting to get raised. And this is in a space where AI has been able to most obviously generate some value. As you move into areas of business where this is less obvious the questions get harder and harder to answer. And the pricing models are going to change across more, if not all, use cases as the easy money dries up and big-tech starts looking to actually, you know, and this is shocking, generate some money.
Based on the word salad that is generally used by the likes of Sam Altman and Satya Nadella, not to mention the hype being nowhere near the reality, I don’t think they are ready for this.
I, for one, hope that AI now starts to get some real scrutiny. Not because I think it is a bad or valueless technology, but because the free ride it’s getting is having real world impacts on real people.