OpenAI Considers Dropping Prices

OpenAI is considering cutting its token prices. Sam Altman said that AI costs have become a real problem for enterprise customers, and a price cut would be one way to address that. This comes as OpenAI and Anthropic both just filed for IPOs, at $852 billion and $965 billion valuations respectively. So we have the market leaders in the most hyped technology of our generation, valued like monopolies, and one of them is contemplating competing on price.

When I studied economics in university, I learned about the Bertrand paradox: when firms sell undifferentiated products and compete on price, you don't need dozens of competitors to destroy profits. Two is enough. Each firm undercuts the other until price falls to marginal cost and economic profit hits zero. A duopoly can produce the same outcome as perfect competition.

Gas stations are the everyday example. Gasoline is fungible, prices are posted on giant signs, and switching costs you five minutes of driving. Margins on fuel are famously a few cents per litre. The real money is made inside the convenience store.

The key word in all of this is "undifferentiated." We now have at least five frontier lab: OpenAI, Anthropic, Google, xAI, and Meta; each building its own model and suite of AI tools. But when you really think about it, it's genuinely hard to tell the models apart in any way that matters. No single model is so superior at one thing that you'd reach for it specifically. And even when one model do pull ahead, the others catch up within months with a competing release. From where most users sit, the models are close to interchangeable, and the existing pricing model makes it prohibitive for users to get more than one subscription.

So what happens next? Do we enter a price war between the frontier labs?

Economists have catalogued the known escapes from Bertrand's trap: differentiation, capacity constraints, and switching costs. I'll take them in turn. Differentiation seems unlikely. I don't see these labs confining a general-purpose technology to a single niche when the whole pitch is that it does everything. Capacity constraints don't hold either; with many billions flowing into data centres and GPU development, the shortage in compute is a temporary issue. That leaves switching costs. And that's the one I think actually has teeth.

This builds on an idea I wrote about last year: the real moat for AI isn't the model, but it's context and memory. By that I mean everything the platform accumulates about you over time: your tone, your drafting preferences, your risk posture, the projects and documents you've fed it, the thousands of past conversations it can draw on. These are extremely hard to consolidate and migrate from one model to the next.

So where are you parking your context?