Ramp’s AI Index for September 2026 shows falling per-employee AI spending among the highest-spending US companies, while usage shifts away from expensive frontier models toward cheaper alternatives.

Financial services company Ramp tracks monthly US business spending on AI services through its AI Index. The September edition shows AI adoption still growing, but at a slower pace. The companies that spend the most on AI spent less in August than they did the month before.

In August, 43.8 percent of US companies paid for Anthropic services, up 0.34 percentage points. OpenAI gained just 0.09 percentage points to reach 39.8 percent. Technical sectors like IT and finance continue to lead adoption.

The biggest spenders are cutting back

Spending by the top 1 percent of companies matters most for model providers because those firms drive the bulk of enterprise revenue. Median per-employee spending in that group fell 9.7 percent in August to $7,205. Because the top 1 percent is a small pool of firms, this estimate tends to be more volatile than other segments and may be revised later, according to the report.

Some of the drop is likely seasonal, since many engineers take vacation in August, according to Ramp chief economist Ara Kharazian. Two other forces are also pushing spending down across industries: falling token prices and a steady migration toward cheaper models.

Token prices keep falling and companies are trading down

The effective price per million tokens has dropped 41 percent since its March 2026 peak to $0.68, according to Ramp. Both OpenAI and Anthropic have announced further price cuts recently. Usage volume is growing, but Kharazian says it may not grow fast enough to offset the price decline.

Most of the volume growth is coming from cheaper standard models like GPT-5.6 Terra and Claude’s Sonnet series. Frontier models like Opus, Fable, and Sol held a 45 percent share of all tokens consumed in early September, down from 53 percent at the start of August. Companies are putting internal policies in place that restrict use of expensive frontier models, Kharazian says, because standard models are increasingly seen as good enough while costing significantly less.

Open-weight models still aren’t moving the needle

The shift toward cheaper models isn’t being driven by open-weight or Chinese alternatives. Only 6.4 percent of AI-using companies on the Ramp platform run open-weight models, and across all companies, that figure drops to 3.6 percent. Since Ramp measures usage through routing platforms that also offer access to closed models, actual open-source adoption is likely even lower.

Last month, Kharazian flagged what he called “Cracks in the AI Thesis” in his report, pointing to weak Fable 5 adoption and the shift toward cheaper models as warning signs for providers. Ramp’s data only covers a slice of the overall market. Things should get clearer when Anthropic files for its IPO, reportedly planned for October.