A recent report indicates that one company’s use of artificial intelligence spiraled out of control, resulting in a staggering $500 million bill.
A consultant is expressing concerns about what this might signal for businesses in the near future. Integrating AI could be more costly than initially thought.
According to a consultant who spoke to Axios, one of their clients racked up about $500 million in AI-related expenses in just one month. This information was shared by an anonymous source.
The high costs stemmed from the company not implementing any limitations on employees who had access to AI licenses for its large language model, known as Claude.
Claude’s provider, Anthropic, has a pricing structure that charges up to $25 per million tokens. While this may sound reasonable, it’s important to note that one token is roughly equivalent to four characters of written English, or about “0.75 words,” as stated on their site.
And yes, punctuation counts towards that total.
When it comes to documents, token consumption can add up quickly. For instance, a PDF can cost around 125,000 tokens, a lengthy document about 25,000 tokens, and a standard web page could be about 2,500 tokens.
The cost for images through Claude Vision is calculated using a specific formula based on their pixel dimensions. For example, a YouTube thumbnail sized 1280 x 720 pixels would incur a cost of approximately 1,229 tokens. Creating roughly 1,000 average-sized images might amount to just under $5—a figure that’s believed to be higher due to workforce scale and how well Claude understands coding.
This unnamed company, which is identified as a U.S. corporation on LinkedIn, reportedly allowed employees unrestricted access with no spending caps or limits.
Without any safeguards, companies generating documents, web pages, or using AI for programming could inadvertently spend tens of thousands of dollars.
A chief technology officer recounted to Axios that staff were employing AI for even the most trivial tasks, like checking the weather. They suggested that the token pricing model isn’t as straightforward as it might seem and certainly isn’t an “all you can eat” plan.
