Live with Tim O’Reilly: A Conversation with Independent Researcher Anjali Shrivastava

A few months ago, Cursor took some heat for updates the company made to its Pro plan, which in effect raised prices for power users. Cursor justified the changes by pointing to growing token use—Anthropic rolled out rate limits at the same time, leading many to speculate that Cursor was just passing along the new costs to its customers. However, researcher and data scientist Anjali Shrivastava thinks that this narrative “gets the causality backwards,” explaining: “Cursor’s users generated the extreme cost tail, because they had no usage limits in place; Anthropic’s response was a downstream response, and Cursor’s price hikes were forced re-pricings of the tail risk.” But her aim isn’t to apportion blame for higher prices. It's to demonstrate that the pricing paradigm for AI is completely broken.

Anjali’s working on a multipart thesis on the unit economics of frontier AI labs and beyond, and she’s shared some of her work in progress with me. At the heart of AI’s pricing dilemma, she notes, is the humble token. We think about the token as an “atomic, static” unit of cost (a legacy from cloud APIs), but in fact token cost is highly variable—which makes it extremely difficult to set a subscription price that doesn’t bring about financial calamity as usage skyrockets. I’ve asked Anjali to join me on this episode of Live with Tim O’Reilly to discuss her work and trace the beginnings of an answer to the question animating it: “How do you design a pricing model that survives margin squeeze [and] fat-tailed costs and doesn’t lead to surprise service degradation?”

We’ll cover the problems with current pricing models from the perspectives of both the model providers and the third-party vendors integrating them into their products. We’ll also examine where this leaves users, who have limited insight into these pricing models and, as Anjali points out, “no incentive to optimize their requests to minimize compute.” And of course we’ll take your questions and use them to prompt further discussion.