American businesses used more AI in August and September and paid almost the same for it. Spending on tokens, the units AI is sold by, grew 1.5% across the two months, after growing 42% in July alone. The amount of work handed to AI kept rising, and the price of each piece of it kept falling.
Three things pull the bill down while the work goes up
The first is list prices. OpenAI cut the price of its GPT-6 Sol and Luna models in half in September, and Anthropic answered with cuts of its own. The average price a business pays for a million tokens has fallen from $1.15 in March to 68 cents.
The second is discounts. The AI companies offer enterprise contracts meant to lock customers in, but buyers are signing them without promising to spend more than they do now, and they still get money off the sticker price. The buyer is the one with leverage.
The third is that customers get better at using AI. A company a year into it sets defaults that stop an expensive model running for hours on a long task, gives each employee a budget, and sends simple jobs to cheaper models.
A lot of business AI work is a choice from a short list: is this email spam, does this invoice line match the purchase order, which team should get this ticket. A new model called Jev answers only from a fixed list of choices, and in one set of tests it processed batches of emails for a few cents per thousand. Move that work off a chat model and the volume stays the same while the bill shrinks.
Why the makers feel it and most buyers do not
The spending that makes AI companies look enormous comes from very few customers. The top 1% of businesses spend close to $10,000 per employee each month, mostly on coding agents, and they account for about 80% of what the sampled businesses pay OpenAI and Anthropic. The median business spends $12 to $15 per employee per month.
So when the heaviest users trim their spending even a little, the makers' revenue stalls, however many more companies start paying. A software firm can find $10,000 a month of work per engineer for coding agents. A doctor's office or a car dealership cannot, so the typical customer will not grow into that number. Anthropic's own filing shows the squeeze: price cuts and discounts took about $600 million off its third-quarter revenue growth.
What a normal company gets from this
Take a company with 200 people at the median. It spends about $2,400 to $3,000 a month on AI. If the same work keeps getting cheaper at the pace it has since March, down 41%, it costs about $1,400 to $1,800. That saving arrives without anyone changing a setting, and it can pay for doing more.
Cheap decisions also make jobs affordable that were never worth doing. A distributor that wants every one of its 20,000 monthly invoice lines checked against the purchase order could do it for around a dollar a month at those prices, where a person would need weeks.
After the fiber glut, the cheap capacity found uses that had been too expensive before. The AI companies are counting on the same thing, and some expect personal agents to become the next coding. That may happen. For a buyer it matters less than the plain fact that the price of each task is falling.
So when a salesperson asks for a commitment to a higher volume than you use now, you do not have to give one. Buyers who can switch providers are getting discounts without making that promise.