Open and closed AI models are on different exponentials

Nathan Lambert argues that open and closed AI are not converging onto one curve. They are running on two different exponentials and will produce two different kinds of winners. Closed frontier labs, he writes, look like a new mix of Apple and Microsoft: integrated stacks selling premium intelligence to people whose work justifies high prices. Open models will not catch up at the top of the curve, but they will absorb a growing share of inference where "good enough" is the bar.

The hinge of the argument is willingness to pay. Lambert estimates that a coding-agent user already gets enough output from a frontier model to make $2,000 a month rational, and he sees that ceiling rising. Anthropic and OpenAI are moving into that segment as platforms, not just APIs, with their own clients, their own SDKs, and increasingly their own opinions about how an agent should be built. He projects the most successful frontier labs into the $2 to $10 trillion range over five to ten years.

Open models, meanwhile, win the long tail. As soon as a use case clears its quality threshold, it tends to stop upgrading, which pushes pricing toward commodity. The implication for builders, Lambert says, is that intelligent routing between open and closed models stops being an optimization. It becomes infrastructure.

Why it matters

If you ship a product with an LLM in it, treat model routing as a first-class design decision now, before pricing pressure makes it urgent. The choice of which calls go to a frontier model and which go to an open one will soon shape your gross margin more than your prompt engineering does.