Ramp data shows AI token prices down 41%, a 'crack' in the boom

The AI infrastructure boom has rested on a core assumption: that demand for frontier models would keep token prices high and cover the massive debts incurred building data centers. Ramp's data challenges that assumption directly, with token prices down 41% since March and usage shifting away from frontier models toward cheaper mid-tier options. OpenAI's 80% price cut on GPT-5.6 Luna — now cheaper than some Chinese open-source models — and Anthropic's own cuts show the labs themselves are driving the decline.

Ramp's chief economist frames this not as a bubble bursting but as a 'crack in the AI thesis.' The pattern mirrors what Citadel Securities noted in June about a 'bifurcation' between expensive frontier AI and the 'everyday' AI most businesses use. Only 3.6% of Ramp's businesses use open-source or Chinese models, yet those markets are also in a brutal pricing war, making the price decline an international phenomenon.

The stakes are clearest in the bond market: Morgan Stanley flags up to $300 billion in bonds financing neocloud buildouts as vulnerable if prices don't recover. Anthropic's pricing power appears to be eroding as OpenAI takes share, with Anthropic charging nearly double all year but that edge 'wearing down.' Whether the labs can hold prices — or shift to outcome-based pricing as OpenAI's CFO suggests — may determine if the infrastructure debt gets repaid or becomes the first real stress point of the AI boom.