Open Source AI Growth Has Yet to Slow Anthropic’s Momentum

On Monday, Decagon CEO Jesse Zhang shared a thought-provoking perspective in a post titled “Everyone is wrong about open source AI in the enterprise.” His argument focuses on one of the more intriguing contradictions emerging in today’s AI market. According to Zhang, as AI deployments mature, businesses—including his own company—are increasingly shifting workloads to lighter, lower-cost models. Despite that transition, overall spending on premium frontier AI models has remained largely unchanged.

His view offers a different way of understanding the relationship between frontier AI models and open-source alternatives. Rather than seeing them as direct rivals, Zhang argues that they serve different roles within the same development cycle. Expensive frontier models are initially used to validate new AI use cases, while those applications eventually move to more affordable open-source models as they mature for large-scale deployment.

As older and more established workloads transition to lightweight models, entirely new AI applications continue to emerge. According to Zhang, that constant cycle helps explain why spending on frontier models remains relatively stable even as open-source adoption grows.

Although Zhang did not include extensive data to support his argument, available industry metrics appear to align with his observations. Data from Vercel’s AI gateway dashboard shows that DeepSeek has rapidly climbed to the top in token volume over the past week, now processing slightly more than one-third of all tokens flowing through the platform. Z.ai, the company behind the increasingly popular GLM-5.2 model, also rose into fourth place during the same period.

However, a look at overall AI spending on the platform tells a different story. Anthropic continues to account for more than half of the total AI expenditure across Vercel’s infrastructure. While its overall share has dipped slightly over the past month—largely due to changes in Anthropic’s pricing—it still maintains a commanding position in revenue generated from model usage.

A similar trend can be seen on OpenRouter, which reflects a broader, though somewhat less enterprise-focused, segment of the AI market. DeepSeek V4 Flash currently leads in overall usage, processing approximately 5.3 trillion tokens each week. Among frontier models, Anthropic’s Opus 4.8 handles just over 2 trillion tokens per week. Although OpenRouter does not rank models by total revenue, it lists the average price of Opus 4.8 at around $1.37 per million tokens, compared with roughly six cents per million tokens for V4 Flash. Based on those figures, Opus 4.8 is still likely responsible for a significantly larger share of overall AI spending.

Those numbers also do not yet reflect Nvidia’s newest Nemotron model, which is expected to become a major competitor thanks to Nvidia’s extensive industry partnerships and the model’s highly adaptable capabilities.

While these statistics do not conclusively prove Zhang’s theory about the AI development lifecycle, they do indicate that frontier AI providers such as Anthropic have not experienced significant financial pressure from the rapid growth of open-source models—at least for now. One possible explanation is that the market for AI-powered applications is expanding quickly enough for premium models to maintain demand by serving new and experimental deployments. As Zhang summarised, “The frontier labs will keep owning discovery. Open source will increasingly own production.” Another possibility is that many advanced AI tasks remain too complex to be handled entirely by lower-cost alternatives.

Whatever the reason, this emerging two-tier structure—where frontier models dominate innovation and open-source models power mature production workloads—could become a lasting characteristic of the AI industry.

Only last September, discussions centred on the possibility that foundation model developers would eventually become commodity providers, supplying core technology while application-layer companies captured most of the value. Some elements of that prediction have materialised, including the growing use of lightweight models in vertical AI products and the continued stability of businesses built around so-called “GPT wrapper” applications.

At the same time, recent market trends suggest that frontier AI companies have managed to retain the most valuable segment of the ecosystem by continuing to command premium prices for their tokens. Based on current industry dynamics, that advantage does not appear likely to disappear in the near future.