I think Ramp's AI report is way more worth reading than all those model benchmark scores After reading it, I realized the most dangerous signal in AI has emerged: Companies aren't unwilling to spend money—they're just starting to refuse to pay extra for 'smarter' solutions. Ramp just released real AI spending data for U.S. companies in July. Anthropic's enterprise adoption rate has already reached 43.5%, surpassing OpenAI's 39.7%. But after its newest and priciest model, Fable 5, launched for a month, it only accounted for 6% of Anthropic's token consumption and 11.4% of its revenue. The contrast is even starker with GPT-5.6 Sol: it accounts for 25% of OpenAI's token usage and 23% of its model spending. Fable 5 is priced at around $10 per million tokens, nearly double GPT-5.6 Sol's price, but companies are spending only about 75% as much on Fable 5 compared to Sol. So here's the question: Companies are clearly more willing to spend on AI—by July, the top 10% of companies were already burning $650 per person per month, and the top 1% were spending $7,400. But these companies, which understand AI the best and have the highest budgets, are increasingly turning to open-source models and model aggregation platforms. Chinese AI companies' open-source strategies and approaches are starting to impact the revenue of cutting-edge global model companies. In the past, model companies could raise prices by releasing smarter next-gen models. Now, companies are calculating ROI themselves: If a task can be done for $1 and achieve 90 points, it's hard to justify paying $10 long-term for 95 points. Model capabilities keep improving, but token unit prices might become harder and harder to increase. So in the future, what big model companies will truly compete for is who can help enterprises scale their token usage by 10x, 100x. The AI industry might eventually look more and more like cloud computing: total spending continues to skyrocket, but the cost per unit of compute keeps getting cheaper.