Large companies spent noticeably less on AI in August. That is not a rejection of the technology: usage and budgets are still growing in many places. But new figures from U.S. payments company Ramp show that the most lucrative customers in particular are spending less per employee. For providers such as OpenAI and Anthropic, that is an important sign that the next phase of growth cannot come from ever more expensive frontier models alone.
Key takeaways
- Among the top 1 percent of AI-using companies in Ramp’s dataset, spending per employee fell 9.7 percent to $7,205 in August.
- At the same time, effective prices per million tokens fell from a March peak of $1.15 to $0.68, according to Ramp.
- Tool adoption is still increasing, but spending is shifting toward cheaper standard models and more tightly managed budgets.
- For buyers, that creates an opportunity to separate model selection from cost control. For vendors, it raises the pressure to sell value rather than raw computing power.
A decline that needs a closer reading
Ramp analyzes payment data from roughly 70,000 companies. It found that 56 percent of its customers paid for AI products in August, only 0.4 percentage points more than in July. The top end of the dataset is particularly striking: spending per employee among the largest 1 percent of AI users fell from $7,976 to $7,205. That is a substantial monthly move, but not yet a verdict on the whole AI market. Ramp itself points to a possible vacation effect, since many teams are less active in August and consume fewer computing resources.
The sample has limits as well. Ramp serves disproportionately technology-focused companies, and the spending figures are not matched against revenue or productivity gains. The data is therefore better understood as an early signal of commercial use than as a complete picture of the U.S. economy. That is also its value: it shows how real invoices are changing, rather than just intentions in board presentations.
Cheaper tokens are changing model selection
The most important reason for lower spending is straightforward arithmetic. Ramp puts the effective price per million tokens at $0.68 today, down from $1.15 in March. When the same work gets cheaper, the bill can fall even as usage rises. That is good news for companies. For model providers, it means that more requests do not automatically translate into more revenue.
There is also a shift in the models companies choose. According to Ramp, businesses increasingly make standard models such as GPT-5.6 Terra and Anthropic’s Sonnet series their defaults because they are fast enough and less expensive for many tasks. Frontier models remain relevant for complex analysis, programming, or consequential decisions. But they lose share when a lower-cost model provides the same quality. As the dispute over Claude Max and its usage limits already showed, buyers are increasingly judging the economic value of access by reliable capacity, not just by a model name.
Broad use, cautious investment
The spending data fits a more nuanced picture from other surveys. Early in September, the Federal Reserve Bank of New York reported that 61 percent of surveyed service firms and 51 percent of manufacturers use AI in business processes. At the same time, three quarters of service firms and more than 90 percent of manufacturers described their investment as limited or modest. AI has reached broad use, but for most businesses it is not yet a budget line that dominates all of IT planning.
A survey by research firm Futurum adds a different cost problem: 46.9 percent of 1,636 technology decision-makers said they were over their AI budget. That does not contradict Ramp. A company can pay less for individual tokens and still exceed its budget when new applications, consulting, data preparation, or governance are added. The relevant figure is not the invoice for one model but the total of use, integration, and control.
What buyers should manage now
The practical conclusion is not to switch off frontier models across the board. Companies should first measure which tasks really need their additional quality. A cheaper standard model may be enough for summaries, classification, or straightforward assistance; consequential decisions require testing, approvals, and clear accountability. Local or open models can also be an option for some data-sensitive workflows. Ramp still sees them among only a small share of customers in its dataset, however, so they do not yet explain the overall trend.
Competition is becoming more sober. Providers must show that premium models do not merely sound better but deliver measurably better outcomes. Buyers, meanwhile, gain negotiating power when they can substitute models and route workloads deliberately. The spending decline at the top is therefore less a warning against AI than a sign of maturity: the market is starting to distinguish between impressive capability and economically useful capability.