On Monday, Decagon CEO Jesse Zhang put forward a provocative new idea in a post titled “Everyone is wrong about open source AI in the enterprise.” The piece tackles one of the most fascinating paradoxes in today’s AI landscape: as AI deployments mature, they increasingly shift to lighter models, even within his own company. Yet overall spending on expensive, cutting-edge models has barely declined. This offers a fresh perspective on the relationship between frontier and open source models. In Zhang’s view, they aren’t rivals, and the success of open source models isn’t coming at the expense of frontier labs. Instead, they represent two stages of the same lifecycle. Expensive frontier models are used to validate use cases, which are then handed off to cheaper open source alternatives as they become more established. As mature applications move to lighter models, new use cases keep emerging, so spending on frontier models remains largely steady. Zhang doesn’t provide much data to back this up, but the evidence isn’t hard to find.

Vercel’s AI gateway dashboard shows that, in just the past week, DeepSeek has surged into the lead for token volume, now processing just over a third of all tokens flowing through the company’s infrastructure. Z.ai, the lab behind the popular GLM-5.2 model, jumped into a respectable fourth place over the same period. However, if you look at total token spend, Anthropic still accounts for more than half of all AI spending on the platform. While much of the recent shift stems from Anthropic’s own rising prices, its share has dropped only slightly over the past month, not dramatically.

OpenRouter tells a similar story, capturing a much larger but slightly less enterprise-focused slice of the market. DeepSeek V4 Flash leads in overall usage, processing 5.3 trillion tokens weekly. The most popular frontier model, Opus 4.8, handles just over 2 trillion tokens. OpenRouter doesn’t rank models by total spend, but it records the average token cost for Opus 4.8 as roughly 23 times higher than V4 Flash—$1.37 per million tokens compared to just 6 cents. That suggests Opus still captures the lion’s share of spending. These figures don’t even account for the newest arrival, Nvidia’s Nemotron, which is poised to leap to the front thanks to Nvidia’s strong connections and the model’s extreme adaptability. While these numbers don’t fully prove Zhang’s lifecycle theory, they do show that frontier labs like Anthropic aren’t suffering much from the rise of open source—at least not yet.

One explanation is that the market for AI-addressable tasks is growing so quickly that top models maintain their position by dominating early-stage deployments. As Zhang puts it, “The frontier labs will keep owning discovery. Open source will increasingly own production.” Another possibility is that, even as clients migrate to open source, many use cases are so demanding that they can’t be entirely replaced by cheaper alternatives. Either way, this two-tiered economy of models may become a stable feature of the AI landscape. As recently as last September, I was writing about the possibility that foundation labs would end up selling coffee beans to Starbucks—serving as commodity inputs while the application layer reaped the rewards. Some parts of that prediction came true: vertical AI players switched to lighter models, and the economics of “GPT wrapper” startups have remained mostly stable. But we’re also seeing that, token for token, frontier providers have held on to the most desirable part of the marketplace: premium token pricing. And that doesn’t seem likely to change anytime soon.