On Monday, CEO Jesse Zhang of Decagon introduced a thought-provoking theory titled “Everyone is mistaken about open source AI in enterprises.” He highlights a fascinating contradiction within the current AI landscape: while established AI applications are increasingly opting for simpler models, the total expenditure on high-end models remains relatively unchanged, even within his own organization.
Zhang offers a fresh perspective on how frontier and open source models relate. He posits that these models are not direct rivals; rather, the success of open source models does not detract from the achievements of frontier labs. Instead, they represent different stages in a model’s lifecycle, with costly frontier models demonstrating viable use cases that can later be adapted into more affordable open source versions as they evolve.
As established use cases shift towards lighter models, new applications continue to emerge, yet overall spending on frontier models shows minimal decrease.
While Zhang provides limited data to substantiate his claims, additional information is accessible. For instance, Vercel’s AI gateway dashboard indicates that DeepSeek has recently become the top performer in token volume, now processing over one-third of the tokens on the platform. Meanwhile, Z.ai, known for the widely used GLM-5.2 model, has also seen a notable surge in its position.
However, when reviewing the overall token expenditures, Anthropic still accounts for more than half of the total spending on the platform. Although some recent fluctuations are attributed to rising prices from Anthropic, the overall share has only decreased slightly in the past month.
OpenRouter presents a similar narrative, gaining a more substantial but less corporate segment of the market. The DeepSeek V4 Flash model ranks highest in overall usage, handling 5.3 trillion tokens weekly, while the leading frontier model, Opus 4.8, processes just over 2 trillion. Although OpenRouter doesn’t track spending by model, it notes that the average cost of tokens for Opus 4.8 is approximately 23 times higher than that of V4 Flash ($1.37 against 6 cents), indicating that Opus still commands a significant portion of the market spending.
Additionally, the recent introduction of Nvidia’s Nemotron model is likely to shift market dynamics due to Nvidia’s industry ties and the model’s adaptability.
These statistics don’t fully validate Zhang’s observations regarding AI lifecycles. Nonetheless, they suggest that frontier labs such as Anthropic are not currently facing severe setbacks due to the rise of open source alternatives. One possible reason for this resilience is the rapidly expanding market for AI-related tasks, allowing top models to secure their position by dominating initial deployments. As Zhang aptly states, “The frontier labs will continue to drive discovery, while open source will take the lead in production.” Another factor could be the complexity of many use cases that, despite the shift to open source, remain challenging to fully substitute with lower-cost models.
Regardless, this bifurcated model economy may solidify as a stable aspect of the AI landscape.
Just last September, discussions were ongoing about how foundational labs may operate akin to suppliers for larger enterprises — for instance, providing essential components while the application layer gains more significant rewards. Certain aspects of that forecast have materialized: vertical AI sectors have increasingly adopted lighter models, and the financial viability of startups utilizing “GPT wrappers” has remained relatively steady.
Yet, it appears that frontier providers have maintained dominance in the more lucrative segment of the market, securing the premium token pricing. This trend is not expected to change in the near future.
