Two things seem to be shrinking: the window where a lab has a monopoly on frontier intelligence, and the share of AI usage that needs the frontier at all. Which begs the question: why would you even push the frontier?

When ChatGPT came out in November 2022, OpenAI had something nobody else could recreate. It took almost a year and a half until another closed lab called Anthropic caught up with Opus 3. Open-weight models weren't even in the equation. That feels different now. Qwen3.8 caught up with Fable 5 around two months after it came out. The shorter this window gets, the less time a lab has to earn back what it spent.

At the same time, four years ago, if you wanted to be productive with AI, you needed to use the frontier models. Today, many of the same tasks can be completed with smaller and cheaper models, like DeepSeek V4 Flash, which costs roughly 100 times less than Fable 5.

Predictions

  1. 1.1Most tasks done with AI today don’t need frontier intelligence. Consumers will not pay a high price for it, and the most popular products will be heavily subsidized with ads. Essentially the same model as Google Search.
  2. 1.2Google is well positioned to win this category. It has the distribution, the knowledge to build small, fast and reliable models, and the infrastructure to serve them. It literally owns the whole stack: training, serving and the application layer.
  3. 1.3Meta is also well positioned. Arguably product will matter more than the intelligence of the model itself, and Meta is better at product.
  4. 2.1Frontier intelligence will cost a very, very high price to pay back what it cost to train. It will mainly be available to governments and big companies that can turn intelligence directly into a business advantage. It might not even be priced through API costs, since the cost of inference barely matters compared to the value of the result. If a model helps a pharma company find a cure, the price of the tokens is honestly irrelevant. They are paying for the cure. Think mathematical discoveries, new weapons or machines that can go to space.
  5. 2.2OpenAI and Anthropic have the highest chances to win this second category.
  6. 3There will be something in between. Tasks where people need the highest level of intelligence they can get, but cannot afford the frontier price. They are capital constrained, but willing to pay a little. This middle will be served by endless providers using Chinese models through routers like OpenRouter, with providers such as Novita, SiliconFlow, DeepInfra, Fireworks, Together and Workers AI competing on price.

There are a couple of things that could make these predictions wrong.

First, a more intelligent model might actually be cheaper to get the job done. It could need fewer tokens, fewer attempts and less fixing afterwards. I think I’m sometimes seeing this with Astra. So a cheaper model per token isn’t necessarily cheaper per task, and the frontier labs could keep more of that demand than I’m assuming.

Second, how much does catching up depend on being able to learn from the frontier models? Distillation is one way to do that, though it isn’t the only way competitors improve. What if the labs never make their strongest models public? They could use them internally to build the next model, or to train smaller models that are still very intelligent and cheap to run. If those smaller models were better and cheaper than what everyone else could offer, the big labs could capture a lot of the everyday tasks too.

That could become a loop. The big model helps build smaller models. The smaller models make money, alongside paid private access to the big model. That money funds the next big model, which helps build even better smaller models.

If the internal models become better at AI research than even the best human researchers, catching up could become much harder. Competitors might have to spend huge amounts just to reach a level the leading labs already sell cheaply. By then, those labs could be another step ahead.

A possible loop A big model kept private helps train better, cheaper small models. Small models are sold to consumers to make money. Private access to the big model is sold to governments and big corporations, also generating revenue. Revenue funds research and training for the next big model. helps train sell to consumers to make money funds the next model Big model kept private Small models better + cheaper Revenue sell private access to governments + big corporations

Scratchpad