
Be wary of AI costs. You may be spending more than you think and end up with nothing to show for it but an eye-popping invoice.
readers are finding themselves on the sharp, pointy end of AI billing, confused about which costs are covered by what, and surprised by "features" that burn through credit, seemingly without warning.
Startups without the deep pockets of established companies are the first to feel the squeeze.
Sometimes, the approach feels more like an app that urges users to deposit extra cash to keep that streak going than a helpful coding assistant. Take one reader, who told about Claude burning through credits at a ferocious rate after a cache setting was reset in the background.
The cache setting controlled how long Claude "remembered" requests and was set to one hour. It's a useful feature, and it lets Claude reuse material that has already been processed. Once it expires, continuing unfinished work can mean paying to process old material again.
"After hitting my $200 Ultra plan limit," our reader told us, "I bought $250 in additional credits to finish 8 coding threads."
There didn't seem to be much work left to do, so the extra credits should have been plenty. Alas, no, after just under two hours, the entire balance was gone, and no meaningful progress was made.
They dug deeper and found that the newer request had switched back to the five-minute cache "without my action or knowledge." The figures didn't add up either, so he made a request to Anthropic to explain the cache setting change and the cost discrepancy.
"I found myself stuck with its support chatbot, Fin. It described the cache switch behavior as a 'feature' and said it couldn't provide the reconciliation or investigation." The support thread was then closed. Our reader asked again for an explanation. "Fin kept closing the support thread, sometimes just 31 seconds after I asked for human escalation."
asked Anthropic for an explanation, and how users might reach a human when a chatbot comes to the end of its usefulness, but the company has yet to respond.
It's all a bit of a cautionary tale. Our reader said, "The money was gone, the work was unfinished, and I was spending time chasing answers that I needed to spend building my app.
"I wish I'd known this could happen before buying those credits."
Indeed. There are endless stories of how using AI has resulted in unexpected costs. Even celebrity mathematician Dr Hannah Fry found the token costs of sending an agent on an errand a surprise. And then there are the misunderstandings when credits gifted under one program turn out to be inapplicable for another, and support teams play a variant of ping-pong, just with customer requests.
The problem is that burning through credits or running up bills that could wipe out a start-up is all too easy. We've been here before with the cloud, and it appears the lessons need to be learned again.
Our reader complained, "We deserve to understand what we're paying for and to reach a human when the numbers don’t add up."
Caution is advised. Especially when a service appears to need only a few more dollars of credit to finish off that task. ®